内容

<div align="center"> # NEXUM : The Agentic Universe ### *The Definitive Open-Source Directory & Architectural Standard for Autonomous AI Systems (2026 Edition)* <p align="center"> <b>Architecting the Autonomous Frontier • 380+ Curated Agents, Orchestration Frameworks, Sandboxes, Protocols & Design Patterns</b> </p> [![Typing SVG](https://readme-typing-svg.demolab.com?font=Fira+Code&weight=600&size=19&pause=1000&color=0969DA&center=true&vCenter=true&width=840&lines=The+World's+Most+Authoritative+Agentic+AI+Directory;Architecting+the+Future+of+Autonomous+Distributed+Systems;380%2B+Elite+Agents+%7C+Frameworks+%7C+MCP+Protocols;Vetted%2C+Benchmarked+%26+Verified+for+Q3+2026.)](https://git.io/typing-svg) <br/> [![Awesome](https://awesome.re/badge-flat2.svg)](https://awesome.re)  [![Edition](https://img.shields.io/badge/Edition-Q3%202026%20GA-2da44e.svg?style=flat-square&logo=gitbook&logoColor=white)](#01-whats-new-right-now-q3-2026-digest)  [![Catalog Size](https://img.shields.io/badge/Catalog-380%2B%20Verified%20Tools-8250df.svg?style=flat-square&logo=databricks&logoColor=white)](#03-the-agentic-stack)  [![Security Standard](https://img.shields.io/badge/Security-OWASP%20ASI%20Top%2010-cf222e.svg?style=flat-square&logo=owasp&logoColor=white)](#08-guardians-security--governance)  [![License: MIT](https://img.shields.io/badge/License-MIT-0969da.svg?style=flat-square&logo=opensourceinitiative&logoColor=white)](https://opensource.org/licenses/MIT)  [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-1f883d.svg?style=flat-square&logo=github&logoColor=white)](http://makeapullrequest.com)  [![Discussions](https://img.shields.io/badge/Community-Discussions-6f42c1.svg?style=flat-square&logo=githubdiscussions&logoColor=white)](https://github.com/HA2345567/awesome-autonomus-ai-agents/discussions) <br/> [**What's New**](#01-whats-new-right-now-q3-2026-digest)  •  [**Architecture**](#02-the-nexum-architecture)  •  [**The Stack**](#03-the-agentic-stack)  •  [**Design Patterns**](#04-agent-reasoning--design-patterns)  •  [**MCP Registry**](#38-production-mcp-servers--registry)  •  [**Vertical Agents**](#06-vertical--industry-agents)  •  [**Guardians & Safety**](#08-guardians-security--governance)  •  [**Commerce Rails**](#09-agentic-commerce--payments-protocols)  •  [**Contribute**](#14-join-the-vanguard) </div> --- ## 01. What's New Right Now (Q3 2026 Digest) > [!IMPORTANT] > **Ecosystem Velocity Report** — High-priority breakthrough signals across frontier hybrid reasoning, parallel coding IDEs, autonomous sales & cyber defense swarms, and agentic settlement standards. - **Models & Frontier Reasoning** `Claude 3.7 Sonnet` `Anthropic Mythos 5` `OpenAI o3/o4-mini` `xAI Grok 3.5` `DeepSeek V4/R1` Claude 3.7 Sonnet establishes the hybrid standard with dynamically allocated thinking budgets; Anthropic restores Mythos 5 & Fable 5 following compliance reviews; OpenAI o3/o4-mini & GPT-5.5 frontier suites deployed; xAI ships Grok 3.5 live search reasoning; DeepSeek V4 & R1 scale open distillation. → *Explore: [Frontier Models](#31-intelligence-frontier-models)* - **Orchestration & Swarms** `Microsoft MAF 1.0` `Google ADK 2.0` `Swarms Enterprise` `PydanticAI V2` AutoGen & Semantic Kernel unified into Microsoft Agent Framework (MAF 1.0); Google ADK 2.0 graph engine released at Google I/O; Swarms framework powers concurrent multi-agent teams; PydanticAI V2 establishes harness-first composable primitives. → *Explore: [Orchestration Frameworks](#32-orchestration-frameworks)* - **Parallel Coding Agents & IDEs** `Google Antigravity 2.0` `OpenAI Codex CLI` `Cursor 3` `Claude Code` `Void` Google Antigravity 2.0 ships split Editor/Manager parallel workflows; OpenAI Codex CLI (#1 Terminal-Bench 2.1) and Claude Code dominate CLI terminals; Devin Desktop and Cursor 3 Composer 2.5 introduce multi-worktree execution; Void emerges as the premier open-source Cursor alternative. → *Explore: [Engineering Agents](#61-engineering--systems)* - **Autonomous Frontier Sectors** `Cybersecurity & Red-Teaming` `Supply Chain` `Sales & GTM Swarms` Google Big Sleep / Project Naptime pioneers automated zero-day triage; 11x Alice and Artisan Ava deploy autonomous revenue swarms; Project 44 and Flexport automate predictive freight routing. → *Explore: [Vertical Sectors](#06-vertical--industry-agents)* - **Commerce & Settlement Rails** `Google AP2` `Coinbase x402` `Nano XNO` `OpenAI ACP` `Stripe MPP` Google AP2 cryptographic mandates standard, Linux Foundation x402 micropayments protocol, Nano XNO feeless layer-1 settlement, OpenAI ACP, and Stripe MPP process real-world autonomous micro-transactions with zero API keys. → *Explore: [Commerce Protocols](#09-agentic-commerce--payments-protocols)* - **Security & Governance Standards** `OWASP ASI Top 10` `Least-Agency Baseline` `PyRIT` OWASP Top 10 for Agentic Applications (ASI Top 10) adopted as global enterprise baseline addressing Goal Hijacking (ASI01) and Context Poisoning (ASI04) alongside Microsoft PyRIT automated red-teaming. → *Explore: [Guardians & Safety](#08-guardians-security--governance)* --- ## 02. The NEXUM Architecture The NEXUM framework categorizes the autonomous universe into **8 tightly coupled architectural layers** — from human-in-the-loop interaction surfaces down to sandboxed micro-VM runtimes and autonomous financial settlement rails. <div align="center"> ```mermaid graph TD User(["[ Human Operator / User Surface ]"]) subgraph AX ["Layer 1: Agentic Experience (AX)"] BU["Browser & GUI Computer Use"] VO["Sub-80ms Real-Time Voice"] UI["Dynamic Agentic Canvases & Artifacts"] end subgraph STACK ["Layer 2: Core Infrastructure"] FM["Frontier Reasoning Models (Claude 3.7 / GPT-5 / Gemini 3.5 / DeepSeek / Grok)"] AF["Orchestration & Swarm Engines (MAF / LangGraph / ADK / CrewAI / Swarms)"] MA["Micro-Agents & Lightweight SDKs (Smolagents / PydanticAI / Rig)"] SAND["Sandboxes & Micro-VM Runtimes (E2B / Daytona / Modal / Fly.io)"] MEM["Temporal Memory & Knowledge Graphs (Mem0 / Letta / Graphiti / HippoRAG)"] PR["Universal Protocols (MCP / A2A v1.0 / UCP / AG-UI)"] MCP["Production MCP Servers (Postgres / GitHub / Playwright / Kubernetes)"] end subgraph PATTERNS ["Layer 3: Reasoning & Design Patterns"] REACT["ReAct & Reflection Loops"] PLAN["Plan-and-Solve & Tree of Thoughts"] SWARM["Hierarchical Supervisor & Worker Swarms"] ROUT["Dynamic Intent Routing & Self-Assembly"] HITL["Human-in-the-Loop Verification Gateways"] end subgraph VERT ["Layer 4: Vertical & Industry Agents"] DEV["Software Engineering & IDEs"] CYBER["Cybersecurity & Autonomous Defense"] SALES["Sales, GTM & Revenue Swarms"] RES["Deep Research & Synthesis"] OPS["DevOps, SRE & Cloud"] BIZ["Finance & Operations"] CX["Customer Support Swarms"] SUPPLY["Supply Chain & Logistics"] DEFENSE["Defense, Aerospace & Security"] ENERGY["Energy, Utilities & Climate"] INSUR["Insurance & Underwriting"] SCI["Science, BioTech & Materials"] DATA["Data Analytics & SQL Reasoning"] ROB["Physical AI & Humanoid Robotics"] LAW["Legal, Contract & Compliance"] end subgraph OBS ["Layer 5: Evaluation & Observability"] BENCH["Benchmarks (SWE-bench / Terminal-Bench / GAIA / OSWorld)"] CORPUS["Datasets & Golden Trajectories (ToolBench 2 / Mind2Web)"] TRACE["Tracing & Evaluation (LangSmith / Langfuse / Phoenix / DeepEval)"] end subgraph SEC ["Layer 6: Guardians & Governance"] GATE["Security Gateways & Runtime Guardrails"] OWASP["OWASP ASI Top 10 Risk Mitigation"] RED["Automated Red-Teaming (PyRIT / Garak)"] end subgraph ECON ["Layer 7: Operations & Agentic Economy"] PAY["Commerce Settlement (AP2 / x402 / ACP / MPP)"] WORK["No-Code & Visual Orchestration (n8n / Dify / Flowise)"] end User <===> AX AX <===> AF AF <---> PATTERNS AF <---> MA AF <---> SAND AF <---> MEM AF <---> FM PR <---> AF MCP <---> AF AF <===> VERT VERT <---> OBS VERT --- SEC VERT <---> ECON classDef userStyle fill:#0969da,stroke:#58a6ff,stroke-width:2px,color:#ffffff; classDef axStyle fill:#161b22,stroke:#58a6ff,stroke-width:1.5px,color:#ffffff; classDef stackStyle fill:#0d1117,stroke:#388bfd,stroke-width:1.5px,color:#ffffff; classDef patternStyle fill:#1a102f,stroke:#a371f7,stroke-width:1.5px,color:#ffffff; classDef vertStyle fill:#0d1117,stroke:#2ea043,stroke-width:1.5px,color:#ffffff; classDef obsStyle fill:#0d1117,stroke:#56d364,stroke-width:1.5px,color:#ffffff; classDef secStyle fill:#2b0c10,stroke:#f85149,stroke-width:1.5px,color:#ffffff; classDef econStyle fill:#211e05,stroke:#e3b341,stroke-width:1.5px,color:#ffffff; class User userStyle; class BU,VO,UI axStyle; class FM,AF,MA,SAND,MEM,PR,MCP stackStyle; class REACT,PLAN,SWARM,ROUT,HITL patternStyle; class DEV,CYBER,SALES,RES,OPS,BIZ,CX,SUPPLY,DEFENSE,ENERGY,INSUR,SCI,DATA,ROB,LAW vertStyle; class BENCH,CORPUS,TRACE obsStyle; class GATE,OWASP,RED secStyle; class PAY,WORK econStyle; ``` </div> --- ## Table of Contents | Section | Architecture Domain | Detailed Breakdown | | :--- | :--- | :--- | | **[01. What's New](#01-whats-new-right-now-q3-2026-digest)** | Digest | Q3 2026 breakthroughs across frontier reasoning, models, tools, and protocols | | **[02. Architecture](#02-the-nexum-architecture)** | Blueprint | 8-layer structural blueprint of the autonomous ecosystem | | **[03. The Agentic Stack](#03-the-agentic-stack)** | Core Infrastructure | [Frontier Models](#31-intelligence-frontier-models) • [Orchestration](#32-orchestration-frameworks) • [Micro-Agents](#33-micro-agents--lightweight-sdks) • [Sandboxes](#34-agent-sandboxes--execution-runtimes) • [Memory](#35-agent-memory--context-layer) • [Protocols](#36-connectivity-protocols--standards) • [Local & Edge](#37-local--edge-on-device-agents) • [MCP Servers](#38-production-mcp-servers--registry) | | **[04. Design Patterns](#04-agent-reasoning--design-patterns)** | Architecture | [ReAct](#41-react-reasoning--acting) • [Plan-and-Solve](#42-plan-and-solve--tree-of-thoughts) • [Reflexion](#43-reflexion--verbal-reinforcement) • [Hierarchical Swarms](#44-hierarchical-supervisor-swarms) • [Dynamic Routing](#45-dynamic-intent-routing--self-assembly) • [HITL Gateways](#46-human-in-the-loop-hitl-gateways) | | **[05. Agentic Experience (AX)](#05-the-agentic-experience-ax)** | Interaction | [Browser & GUI Computer Use](#51-vision--control-browser--computer--gui-use) • [Voice & Streaming Presence](#52-voice--presence) • [Dynamic Canvases](#53-dynamic-canvases-artifacts--agent-user-interfaces) | | **[06. Vertical & Industry Agents](#06-vertical--industry-agents)** | Applied AI | [Engineering](#61-engineering--systems) • [DevOps/SRE](#62-devops-sre--cloud-infrastructure) • [Deep Research](#63-deep-research--intelligence) • [Business/Finance](#64-business-finance--hr) • [CX Support](#65-customer-experience--support) • [Enterprise Ops](#66-work-management--enterprise-ops) • [Science/Healthcare](#67-science-biotech--materials) • [Data Analytics](#68-data-analytics--mathematics) • [Gaming](#69-virtual-worlds--gaming) • [Robotics](#610-physical-ai-robotics) • [Legal](#611-law-legal--compliance) • [Design](#612-creative-media--design) • [Education](#613-education--adaptive-learning) • [Cybersecurity](#614-cybersecurity--threat-intelligence) • [Supply Chain](#615-supply-chain--autonomous-logistics) • [Sales & GTM](#616-sales-go-to-market-gtm--revenue-swarms) • [Defense & Intelligence](#617-defense-aerospace--national-security) • [Energy & Climate](#618-energy-utilities--climate-tech) • [Insurance & Underwriting](#619-insurance--risk-underwriting) | | **[07. Evaluation & Observability](#07-agent-evaluation-tracing--observability)** | Telemetry | [Benchmarks & Leaderboards](#71-benchmarks--leaderboards) • [Datasets & Trajectory Corpora](#72-datasets-golden-traces--trajectory-corpora) • [Tracing & Telemetry Suites](#73-tracing-telemetry--evaluation-suites) | | **[08. Guardians & Governance](#08-guardians-security--governance)** | Safety | [Security Gateways & Guardrails](#81-security-gateways--guardrails) • [OWASP ASI Top 10 Risk Matrix](#82-owasp-top-10-for-agentic-applications-asi-top-10) • [Automated Red-Teaming](#83-automated-red-teaming--adversarial-probing) | | **[09. Agentic Commerce](#09-agentic-commerce--payments-protocols)** | Monetization | [Commerce Protocols & Wallets](#91-commerce-protocols--wallets) • [Payment Protocol Matrix](#92-agentic-payments-protocol-matrix) | | **[10. Operations & Economy](#10-agent-operations--economy)** | Operations & Billing | [Workflow & Low-Code Engines](#101-workflow--low-code) • [Metering & Sub-Cent Billing](#102-metering--billing) | | **[11. Research Papers & Literature](#11-seminal-research-papers--literature)** | Foundations | Essential reading list of seminal agent papers (2022–2026) | | **[12. Multi-Language Framework Matrix](#12-multi-language-agent-framework-matrix)** | Programming | Python, TypeScript, Go, Rust, C#, Java SDK index | | **[13. The NEXUM Standard](#13-the-nexum-standard)** | Methodology | Lead Designer Vetting Process & Inclusion Criteria | | **[14. Join the Vanguard](#14-join-the-vanguard)** | Community | [Contribution Workflow](#141-contributing-workflow) • [Contributors](#142-contributors) • [Star History](#143-star-history) | --- ## 03. The Agentic Stack *The core infrastructure layer powering the autonomous world.* ### 3.1. Intelligence (Frontier Models) - **[Claude 3.7 Sonnet / Sonnet 5 / Opus 4.8](https://www.anthropic.com/claude)** `Proprietary` `Hybrid-Reasoning` — Anthropic's flagship agentic line featuring Claude 3.7 Sonnet's hybrid architecture (instant response + dynamic thinking budget), Opus 4.8 for deep reasoning, Sonnet 5 for fast workflows, and Haiku 4.5 for low-latency tasks. - **[Claude Mythos 5 / Claude Fable 5](https://www.anthropic.com/news/fable-mythos-access)** `Proprietary` `High-Security` — Anthropic's pinnacle tier designed for high-assurance autonomous planning; Fable 5 provides added alignment barriers for critical cyber/bio/LLM-R&D domains. - **[GPT-5.4 / GPT-5.5 Family](https://openai.com/index/introducing-gpt-5-4/)** `Proprietary` `Flagship` — OpenAI's frontier reasoning family for professional work, multi-modal coding, computer use, and long-horizon agent execution; GPT-5.5 tops public Terminal-Bench 2.1 via Codex CLI. - **[OpenAI o3 / o4-mini / o1](https://openai.com/)** `Proprietary` `Reasoning` — OpenAI's frontier reasoning models featuring adaptive chain-of-thought compute, deep multi-step code synthesis, and native function calling. - **[Gemini 3.1 / 3.5 Pro & Flash](https://ai.google.dev/models/gemini)** `Proprietary` `Multimodal` — Google's multimodal frontier series with 2M+ context window, native tool execution, agentic coding, and the low-latency Gemini 3.5 Flash tier powering Antigravity 2.0. - **[xAI Grok 3 / Grok 3.5](https://x.ai/)** `Proprietary` `Real-Time Search` — xAI's frontier model suite featuring multi-step chain-of-thought reasoning, live X-stream data ingestion, and high-throughput agent tool execution. - **[DeepSeek V4 / V3.2 / R1](https://api-docs.deepseek.com/quick_start/pricing/)** `Open Weights` `Reasoning` — DeepSeek's open reasoning and coding model line, offering cost-efficient frontier intelligence, open weights distillation, and native agentic function calling. - **[Qwen 3.0 Max / Qwen 2.5-Coder](https://github.com/QwenLM/Qwen)** `Open Weights` `Coding Standard` — Alibaba's open-weight model family with native MCP integration, 1M context, and top open-source SWE-bench scores. - **[Meta Llama 4.5 Agentic / Llama 3.3 70B](https://llama.meta.com)** `Open Weights` — Meta's open-weight foundation model family optimized for local agent deployment, tool execution, and multi-agent coordination. - **[Kimi K2.5](https://github.com/MoonshotAI)** `Open Weights` — Moonshot AI's open-source base model, serving as the foundation under Cursor's in-house Composer 2.5 agent after heavy RL post-training. - **[Mistral NeMo Agentic 2 / Large 3](https://mistral.ai/)** `Open Weights` `Tool-Optimized` — Mistral AI's agent-optimized open model series engineered for high-throughput tool calling, structured outputs, and complex MCP action execution. - **[Cohere Command R+ (Agentic)](https://cohere.com/command)** `Commercial` `Enterprise-RAG` — Scalable enterprise model purpose-built for multi-step tool use, grounded RAG citations, and automated business workflows. - **[GLM-4-Voice / GLM-4-Plus](https://github.com/THUDM/GLM-4)** `Open Weights` — Zhipu AI's bilingual open model family with direct end-to-end multimodal voice processing and tool execution loops. ### 3.2. Orchestration (Frameworks) - **[Microsoft Agent Framework (MAF) 1.0](https://github.com/microsoft/agent-framework)** `MIT` `GA Release` — Microsoft's official successor to AutoGen and Semantic Kernel, reaching GA across Python and .NET with declarative YAML config, native MCP, and A2A integration. - **[LangGraph 1.0](https://github.com/langchain-ai/langgraph)** `MIT` `Enterprise Standard` — The enterprise standard for stateful, cyclic multi-agent graphs, featuring MCP tools as first-class streaming graph nodes and time-travel debugging. - **[Google Agent Development Kit (ADK) 2.0](https://github.com/google/adk)** `Apache 2.0` `Graph Engine` — Google's code-first agent toolkit rebuilt around a graph-based execution engine with deep Gemini, Cloud Run, and A2A integration. - **[Claude Agent SDK](https://pypi.org/project/claude-agent-sdk/)** `MIT` `Anthropic Official` — Anthropic's official SDK for building custom agents on the Claude Code harness, using MCP as its primary tool contract. - **[CrewAI 1.14](https://github.com/joaomdmoura/crewAI)** `MIT` `54k Stars` — Role-based multi-agent collaboration framework with built-in telemetry, enterprise flows, and native MCP/A2A swarming support. - **[Swarms](https://github.com/kyegomez/swarms)** `MIT` `18k Stars` — Enterprise-grade multi-agent orchestration framework for building, deploying, and managing concurrent swarms across heterogeneous models. - **[PydanticAI V2](https://github.com/pydantic/pydantic-ai)** `MIT` `Type-Safe` — Type-safe Python agent framework with a "harness-first" redesign bundling tools, hooks, instructions, and model settings into composable capability primitives. - **[AutoGen (v0.4+ Event-Driven)](https://github.com/microsoft/autogen)** `MIT` `Distributed Actors` — Microsoft's redesigned asynchronous, event-driven multi-agent framework featuring distributed agent actors, pluggable memory, and scalable swarms. - **[Semantic Kernel](https://github.com/microsoft/semantic-kernel)** `MIT` `Multi-Language` — Microsoft's enterprise-grade SDK for integrating AI services into C#, Python, and Java applications with rich plugin architecture and agent orchestration. - **[Mastra](https://github.com/mastra-ai/mastra)** `Apache 2.0` `TypeScript Native` — TypeScript-native agent framework tailored for full-stack JavaScript and Node.js teams with built-in RAG and eval. - **[MetaGPT](https://github.com/geekan/MetaGPT)** `MIT` `46k Stars` — Multi-agent framework that assigns Standard Operating Procedures (SOPs) to simulated software company roles (PM, Architect, Engineer, QA). - **[ChatDev](https://github.com/OpenBMB/ChatDev)** `Apache 2.0` `26k Stars` — Communicative agent ecosystem simulating a virtual software development company to build applications from scratch through collaborative dialogue. - **[LlamaIndex Workflows 1.0](https://github.com/run-llama/llama_index)** `Apache 2.0` `Event-Driven RAG` — Event-driven workflow engine from LlamaIndex for RAG-heavy and data-centric multi-agent pipelines with native MCP support. - **[AG2](https://github.com/ag2ai/ag2)** `Apache 2.0` — Community-driven continuation of original AutoGen v0.2, maintained by former creators. - **[Camel-AI](https://github.com/camel-ai/camel)** `MIT` — Pioneer multi-agent framework exploring communicational agent behavior, role-playing, and cooperative task execution. - **[Agno (formerly Phidata)](https://github.com/agno-agi/agno)** `MIT` `High-Performance` — High-performance framework for lightweight, multimodal agents with vector memory and fast reasoning loops. - **[Microsoft Magentic-One](https://github.com/microsoft/autogen/tree/main/python/packages/autogen-magentic-one)** `MIT` — Generalist multi-agent system from Microsoft Research capable of solving complex multi-step web and file-based tasks. - **[SuperAGI](https://github.com/TransformerOptimus/SuperAGI)** `Apache 2.0` — Open-source developer-first autonomous AI agent framework designed to build, manage, and run autonomous agents with GUI dashboard management and multi-vector memory. - **[AutoGen Studio](https://github.com/microsoft/autogen/tree/main/python/packages/autogen-studio)** `MIT` `Visual UI` — Interactive low-code UI for rapid multi-agent prototyping, visual skill creation, and session debugging. ### 3.3. Micro-Agents & Lightweight SDKs - **[Smolagents](https://github.com/huggingface/smolagents)** `Apache 2.0` `44k Stars` — Hugging Face's ultra-lightweight library for code-first agents where actions are clean Python code blocks. - **[ControlFlow](https://github.com/PrefectHQ/ControlFlow)** `Apache 2.0` `Task-Centric` — Prefect's task-centric, Pythonic agent framework designed to weave autonomous agents into structured, observable workflows. - **[Rig](https://github.com/0xPlaygrounds/rig)** `MIT` `Rust SDK` — Modular, type-safe Rust library for building scalable, high-performance LLM-powered applications and agents. - **[Goose](https://github.com/block/goose)** `Apache 2.0` — Block's open-source extensible AI agent that runs in your terminal to automate shell scripts, code edits, and environment operations. - **[Micro-Agent](https://github.com/builderio/micro-agent)** `MIT` — Builder.io's lightweight agent that writes code, executes tests in a sandbox, and iteratively fixes errors until tests pass. - **[BabyAGI 2.0](https://github.com/yoheinakajima/babyagi)** `MIT` — The evolution of the original autonomous task planner into a minimalist, modular micro-agent framework. - **[OpenBB Agents](https://github.com/OpenBB-finance/OpenBBTerminal)** `Apache 2.0` `Financial SDK` — Specialized lightweight financial research micro-agents grounded in market data endpoints. ### 3.4. Agent Sandboxes & Execution Runtimes - **[E2B (Code Interpreter Sandboxes)](https://github.com/e2b-dev/E2B)** `Apache 2.0` `22k Stars` — Secure cloud micro-VM sandboxes used by Devin, Cursor, and Perplexity for running agent code safely isolated. - **[Daytona](https://github.com/daytonaio/daytona)** `Apache 2.0` — Open-source development environment manager and agent sandbox runtime enabling instant isolated workspace provisioning. - **[Modal Micro-VM Runtimes](https://modal.com/)** `Commercial` `Serverless GPU` — Serverless infrastructure optimized for ultra-low latency agent sandboxing, code execution, and GPU workloads. - **[Fly.io MicroVM Machines](https://fly.io/docs/machines/)** `Commercial` `Global Edge` — Ephemeral micro-VM infrastructure for provisioning isolated, sub-second execution sandboxes across global edge locations. - **[Runloop](https://www.runloop.ai/)** `Commercial` `Managed Sandbox` — Cloud platform providing managed developer environments and sandboxes optimized for AI coding agents and benchmark evaluation. - **[Firecrawl Extract Engine](https://github.com/mendableai/firecrawl)** `AGPL-3.0` `Web Sandbox` — Turn-key web scraping and sandbox extraction engine that turns any URL into LLM-ready markdown for autonomous agents. - **[Docker Agent Sandbox](https://www.docker.com/)** `Apache 2.0` — Containerized execution environments optimized for isolating local and enterprise agent file/code operations. - **[Cloudflare Sandboxed Workers](https://developers.cloudflare.com/)** `Commercial` — Edge-based secure compute isolation with native headless browser rendering and sub-millisecond cold starts. ### 3.5. Agent Memory & Context Layer - **[Mem0](https://github.com/mem0ai/mem0)** `Apache 2.0` `28k Stars` — The intelligent memory layer for AI agents; provides persistent user, session, and domain memory across agentic interactions. - **[Letta (formerly MemGPT)](https://github.com/letta-ai/letta)** `Apache 2.0` `Stateful OS` — Stateful AI agent framework managing multi-tier memory (RAM/Disk) like an operating system for long-term retention. - **[Zep / Graphiti](https://github.com/getzep/graphiti)** `Apache 2.0` `Temporal Graph` — Temporal knowledge graph engine that dynamically extracts and updates episodic memory for autonomous agents. - **[HippoRAG](https://github.com/OSU-NLP-Group/HippoRAG)** `Apache 2.0` `Hippocampal Memory` — Neurobiologically inspired long-term memory indexing framework enabling continuous knowledge integration and multi-hop reasoning. - **[Khoj](https://github.com/khoj-ai/khoj)** `AGPL-3.0` `20k Stars` — Open-source personal AI second brain with offline episodic memory, search indexation across documents, and custom agent creation. - **[Supermemory](https://github.com/supermemoryai/supermemory)** `MIT` — Open-source build-your-own memory engine for AI agents with document vectorization and temporal search. - **[Cognee](https://github.com/topobyte/cognee)** `Apache 2.0` — Memory management and graph-RAG pipeline for LLMs and agents, organizing unstructured context into searchable knowledge graphs. - **[VectorShift Memory Engine](https://www.vectorshift.ai/)** `Commercial` — Enterprise memory and knowledge retrieval infrastructure for autonomous multi-step pipelines. ### 3.6. Connectivity (Protocols & Standards) - **[Model Context Protocol (MCP)](https://modelcontextprotocol.io)** `Open Standard` `Anthropic` — The universal open standard ("USB-C for AI") connecting agents to external data sources, tools, and environments; primary tool contract for Claude Agent SDK and MAF 1.0. - **[Agent2Agent (A2A) v1.0](https://developers.google.com)** `Open Standard` `Google` — Open protocol enabling heterogeneous AI agents to negotiate tasks and exchange data; backed by 150+ organizations with production SDKs in Python, JS, Java, Go, and .NET. - **[Agent Network Protocol (ANP)](https://agent-network-protocol.com)** `Open Standard` — Decentralized communication standard for P2P inter-agent messaging and discovery. - **[Universal Commerce Protocol (UCP)](https://developers.google.com/shopping)** `Open Standard` `Google` — Protocol for autonomous buy-and-sell agents to execute commercial transactions across digital storefronts. - **[MCP-A2A Interoperability Bridge v1.2](https://modelcontextprotocol.io)** `Open Standard` — Unified bridge protocol enabling Anthropic MCP tool contexts and Google A2A inter-agent negotiation to seamlessly hand off tasks across agent networks. - **[AG-UI Protocol](https://github.com/ag-ui-protocol/ag-ui)** `Open Standard` — Agent-to-User Interface streaming protocol establishing bidirectional canvas updates, state synchronization, and inline approval cards. - **[OpenAI AgentKit](https://github.com/openai/agent-kit)** `Open Source` — Cryptographic action authorization and wallet interface standard allowing agents to securely interact with third-party Web APIs and on-chain systems. - **[Open-RPC for Agents](https://open-rpc.org/)** `Open Standard` — JSON-RPC-based interface specification for strongly typed agent tool discovery, validation, and contract calling. - **[W3C DID Agent Identity Spec](https://www.w3.org/TR/did-core/)** `Open Standard` — Decentralized Identifiers and verifiable credentials establishing immutable identity and cryptographic accountability for autonomous software agents. ### 3.7. Local & Edge (On-Device Agents) - **[Ollama](https://github.com/ollama/ollama)** `MIT` `120k+ Stars` — The premier local engine for running quantized agent models offline on consumer hardware with zero data leakage. - **[LM Studio Local Server](https://lmstudio.ai/)** `Proprietary / Free` — Desktop local LLM runtime exposing OpenAI-compatible REST endpoints with GPU acceleration and native tool-calling agent capabilities. - **[vLLM Engine](https://github.com/vllm-project/vllm)** `Apache 2.0` `High-Throughput` — High-throughput, memory-efficient LLM serving engine powered by PagedAttention, optimized for concurrent multi-agent tool calls. - **[Apple CoreML Agent Engine](https://developer.apple.com)** `Proprietary` — On-device framework allowing agents to access local file systems, native apps, and Apple Silicon neural engines securely. - **[Exo](https://github.com/exo-explore/exo)** `GPL-3.0` `18k Stars` — Run frontier-class open LLMs locally by pooling consumer devices (MacBooks, PCs, iPhones) into a distributed peer-to-peer AI cluster. - **[LocalAI](https://github.com/mudler/LocalAI)** `MIT` `28k Stars` — Free, open-source OpenAI drop-in replacement for local, self-hosted AI agent inference on CPU and GPU. - **[Jan](https://github.com/janhq/jan)** `AGPL-3.0` — Open-source local desktop AI assistant supporting local tool execution, extension plugins, and offline model runtimes. - **[Orkas](https://github.com/Orkas-AI/Orkas)** `MIT` — Open-source desktop workspace that coordinates coding agents in parallel or sequence and supports local OpenAI-compatible model endpoints. - **[Llama.cpp Agent Server](https://github.com/ggerganov/llama.cpp)** `MIT` — High-efficiency C/C++ inference engine running quantized models with low latency on edge devices and servers. ### 3.8. Production MCP Servers & Registry *Production-ready Model Context Protocol (MCP) servers enabling seamless tool augmentation.* - **[GitHub MCP Server](https://github.com/modelcontextprotocol/servers/tree/main/src/github)** `MIT` — Official MCP server enabling agents to search repositories, inspect pull requests, create issues, and review diffs. - **[PostgreSQL / MySQL MCP Server](https://github.com/modelcontextprotocol/servers/tree/main/src/postgres)** `MIT` — Secure, read/write database inspection server providing schema introspection and parameterized SQL execution. - **[Playwright Browser MCP Server](https://github.com/modelcontextprotocol/servers/tree/main/src/puppeteer)** `MIT` — Headless and visible browser automation server allowing agents to click, scrape, take screenshots, and navigate SPAs. - **[Filesystem MCP Server](https://github.com/modelcontextprotocol/servers/tree/main/src/filesystem)** `MIT` — Sandboxed local directory search, file reading, and atomic file replacement with strict root path boundaries. - **[Brave Search MCP Server](https://github.com/modelcontextprotocol/servers/tree/main/src/brave-search)** `MIT` — Real-time web index querying and citation retrieval for grounded research agents. - **[Slack / Discord MCP Server](https://github.com/modelcontextprotocol/servers/tree/main/src/slack)** `MIT` — Conversational channel integration for dispatching notifications, thread replies, and operational alerts. - **[Docker / Kubernetes MCP Server](https://github.com/modelcontextprotocol/servers)** `Apache 2.0` — Cluster state inspection, container lifecycle management, and log telemetry streaming. - **[Sequential Thinking & Memory MCP](https://github.com/modelcontextprotocol/servers/tree/main/src/sequentialthinking)** `MIT` — Dynamic chain-of-thought scratchpad and persistent context memory retention for multi-step agent reasoning. - **[AWS Cloud MCP Server](https://github.com/awslabs/mcp-server-aws)** `Apache 2.0` — AWS resource provisioning, CloudWatch metric inspection, and IAM policy diagnostics. [Back to Top](#nexum--the-agentic-universe) --- ## 04. Agent Reasoning & Design Patterns *The architectural paradigms governing how autonomous agents think, plan, and collaborate.* ### 4.1. ReAct (Reasoning + Acting) - **Concept**: The agent alternates between an explicit "Thought" reasoning trace and an "Action" tool execution step, observing the environment's response before planning the subsequent step. - **Best Suited For**: Dynamic multi-tool environments, API orchestration, and real-time debugging. - **Key Reference**: *Yao et al. (2022) — ReAct: Synergizing Reasoning and Acting in Language Models.* ### 4.2. Plan-and-Solve & Tree of Thoughts (ToT) - **Concept**: Decouples the strategic macro-planning phase from individual execution tasks. The agent builds a Directed Acyclic Graph (DAG) of sub-goals and explores multiple branching reasoning paths simultaneously. - **Best Suited For**: Complex mathematical proofs, large multi-file codebase refactorings, and multi-day research syntheses. ### 4.3. Reflexion & Verbal Reinforcement - **Concept**: After receiving environment feedback or test failures, the agent generates a natural-language self-critique ("verbal gradient") stored in episodic memory to prevent repeating the same mistake on the next iteration. - **Best Suited For**: Code generation, unit test self-repair, and iterative simulation. - **Key Reference**: *Shinn et al. (2023) — Reflexion: Language Agents with Verbal Reinforcement Learning.* ### 4.4. Hierarchical Supervisor Swarms - **Concept**: A primary Supervisor agent receives user intent, decomposes the problem, and delegates discrete tasks to specialized Worker agents (e.g. Coder, Reviewer, Tester), aggregating their outputs into a cohesive result. - **Best Suited For**: Enterprise workflows, software company simulations (MetaGPT / ChatDev / Swarms), and cross-functional task swarms. ### 4.5. Dynamic Intent Routing & Self-Assembly - **Concept**: The gateway agent dynamically analyzes incoming multi-modal requests, selects the optimal downstream subagent/model specialty on the fly, and injects runtime context-specific tool schemas without static hardcoding. - **Best Suited For**: Open-ended conversational assistants, multi-tenant enterprise gateways, and adaptive customer service swarms. ### 4.6. Human-in-the-Loop (HITL) Gateways - **Concept**: Cryptographic and procedural checkpoints where autonomous agents pause execution to request explicit human authorization before executing irreversible actions (financial transactions, production deployments, database writes). - **Best Suited For**: Agentic commerce (AP2), SOC incident mitigations, and regulated clinical environments. [Back to Top](#nexum--the-agentic-universe) --- ## 05. The Agentic Experience (AX) *How humans control, monitor, and collaborate alongside autonomous agents.* ### 5.1. Vision & Control (Browser / Computer / GUI Use) - **[Browser-Use](https://github.com/browser-use/browser-use)** `MIT` `80k Stars` — The leading open-source library for controlling web browsers via LLM agents; enables automated form filling, scraping, and SaaS interaction. - **[Stagehand](https://github.com/browserbase/stagehand)** `MIT` — Browserbase's AI-first browser automation SDK built on Playwright with natural language actions, extraction, and observation. - **[Skyvern](https://github.com/Skyvern-AI/skyvern)** `Apache 2.0` `11k Stars` — Open-source vision + LLM browser automation agent capable of navigating complex websites without brittle CSS selectors. - **[UI-Tars](https://github.com/bytedance/ui-tars)** `Open Source` — ByteDance's GUI agent model trained for direct vision-based computer and web interaction across Windows, macOS, and mobile. - **[Claude Computer Use / Claude in Chrome](https://claude.com)** `Proprietary` — Direct visual screen perception and mouse/keyboard action execution, available as a dedicated browser agent extension. - **[OpenAI Operator](https://openai.com)** `Proprietary` — Autonomous browser agent capable of executing complex web tasks like travel booking, shopping, and research. - **[Cline / Roo Code](https://github.com/cline/cline)** `MIT` `67k Stars` — VS Code extension agent featuring full autonomous terminal, file edit, and browser preview loops. - **[Open Interpreter](https://github.com/KillianLucas/open-interpreter)** `MIT` `68k Stars` — Natural language command-line agent executing code locally across Python, JS, Shell, and native OS apps. - **[Midscene.js](https://github.com/web-infra-dev/midscene)** `MIT` — Multimodal AI-powered UI automation and visual testing tool for web apps that executes natural language steps via visual grounding. - **[LaVague](https://github.com/lavague-ai/LaVague)** `Apache 2.0` — Large Action Model (LAM) framework generating executable browser automation code directly from visual web pages and natural language goals. - **[Playwright AI Agent](https://github.com/microsoft/playwright)** `Apache 2.0` — Microsoft's autonomous end-to-end web testing and browser automation harness powered by vision-language agents. - **[Agent S (OpenOS)](https://github.com/simular-ai/Agent-S)** `Apache 2.0` — Open-source GUI agent designed to autonomously interact with OS desktops, GUI applications, and multi-window environments. - **[Cradle](https://github.com/BAAI-Agents/Cradle)** `Apache 2.0` — General computer control framework enabling multimodal agents to master desktop applications and open-world video games. ### 5.2. Voice & Presence - **[Vapi.ai](https://vapi.ai)** `Commercial` `Sub-100ms Voice` — Developer platform for building sub-100ms ultra-low latency conversational voice agents with emotional nuance. - **[LiveKit Agents](https://github.com/livekit/agents)** `Apache 2.0` `WebRTC Native` — Open-source framework for building real-time, multi-modal voice and video AI agents over WebRTC with sub-200ms latency. - **[Cartesia Sonic](https://cartesia.ai)** `Commercial` `Sub-80ms API` — Sub-80ms voice generation API designed for real-time streaming agentic voice conversations. - **[Retell AI](https://www.retellai.com/)** `Commercial` — Voice AI platform for building human-like conversational telephone and customer interaction agent swarms. - **[Bland AI](https://bland.ai)** `Commercial` — Enterprise platform for building and scaling autonomous phone-call agent swarms for sales and operations. - **[ElevenLabs + IBM watsonx Orchestrate](https://elevenlabs.io/enterprise)** `Commercial` — Enterprise voice integration bringing ElevenLabs high-fidelity voice synthesis into IBM's agentic workflow platform across 70+ languages. - **[Ultravox v0.5](https://github.com/fixie-ai/ultravox)** `Apache 2.0` `Speech-LLM` — Open-source multimodal speech-LLM designed for direct audio-in audio-out conversational agent systems without transcription bottlenecks. - **[Mini-Omni 2](https://github.com/gpt-omni/mini-omni2)** `Apache 2.0` — Open-source multimodal agent capable of simultaneous real-time voice, vision, and text understanding. ### 5.3. Dynamic Canvases, Artifacts & Agent-User Interfaces - **[Google Antigravity Canvas](https://antigravity.google)** `Proprietary` — Multi-agent supervisor surface rendering live executable artifacts, code diff branches, and parallel subagent execution monitors. - **[Claude Artifacts & Cowork Canvas](https://claude.ai)** `Proprietary` — Dynamic interactive side-by-side workspace rendering live React components, SVG diagrams, and document revisions. - **[OpenAI Canvas](https://openai.com)** `Proprietary` — Collaborative canvas environment enabling targeted code edits, text refactorings, and interactive line-by-line debugging. - **[Bolt.new Dynamic Workspace](https://bolt.new)** `Commercial` — In-browser full-stack development canvas running real Node.js runtimes, hot reload, and instant preview. [Back to Top](#nexum--the-agentic-universe) --- ## 06. Vertical & Industry Agents *Domain-specific intelligence deployed across global enterprise sectors.* ### 6.1. Engineering & Systems - **[Claude Code](https://claude.com/code)** `CLI Native` `SWE-bench Top Tier` — Anthropic's terminal-native agent; highest reasoning ceiling in the category (69.2% SWE-bench Pro with Opus 4.8), featuring parallel subagents and background sessions. - **[Google Antigravity 2.0](https://antigravity.google)** `Freemium` `Parallel Agent IDE` — Google's parallel-agent IDE built on the Windsurf acqui-hire; features split Editor View / Manager Surface for spawning and supervising multiple agents, powered by Gemini 3.5 Flash. - **[Cursor 3 (Composer 2.5)](https://cursor.com)** `Freemium` `Parallel Worktrees` — AI-native code editor powered by Composer 2.5 (built on Kimi K2.5), scoring ~79.8% SWE-Bench Multilingual at lower compute costs with parallel worktree execution. - **[OpenAI Codex CLI](https://openai.com/codex)** `Freemium` `#1 Terminal-Bench` — OpenAI's standalone cloud/CLI coding agent, #1 on the public Terminal-Bench 2.1 leaderboard with GPT-5.5. - **[Void](https://github.com/voideditor/void)** `Apache 2.0` `Open-Source IDE` — The open-source AI code editor alternative to Cursor; provides complete control over model endpoints, privacy, and custom agent tools. - **[Devin Desktop (formerly Windsurf)](https://cognition.ai)** `Commercial` — Cognition's relaunch of the Windsurf IDE, integrating the Devin cloud agent and Devin Terminal CLI for autonomous micro-VM runs. - **[Augment Code](https://www.augmentcode.com/)** `Commercial` `Deep Semantic Index` — Enterprise AI coding platform with proprietary deep codebase semantic indexing, instant dependency-aware context, and collaborative multi-file agents. - **[Codeium Windsurf (Cascade)](https://codeium.com/windsurf)** `Freemium` `Flows Engine` — Agentic IDE featuring Cascade real-time flows, multi-file editing, terminal tool execution, and codebase graph grounding. - **[Kiro](https://kiro.dev)** `Commercial` — Spec-driven development IDE with event-driven hooks for structured spec-first code generation. - **[OpenCode](https://github.com/sst/opencode)** `MIT` `180k+ Stars` — SST's open-source, model-agnostic terminal coding agent; the most-starred open-source terminal coding agent for local offline workflows. - **[Trae](https://www.trae.ai/)** `Freemium` — ByteDance's adaptive AI-native IDE integrating context-aware coding agents into workspace workflows. - **[Google Jules](https://jules.google)** `Freemium` — Google's asynchronous GitHub-issue-to-PR coding agent. - **[Replit Agent 4](https://replit.com)** `Freemium` — Autonomous cloud IDE agent that plans, builds, tests, and deploys full-stack applications with parallel branch merging. - **[GitHub Copilot (Cloud Agent)](https://github.com/features/copilot)** `Commercial` — GitHub-native issue-to-PR automation with flexible billing and cloud agent delegation. - **[Aider](https://github.com/paul-gauthier/aider)** `MIT` `30k Stars` — Command-line pair programmer with git integration and automated commit formatting. - **[SWE-agent](https://github.com/princeton-nlp/SWE-agent)** `MIT` — Princeton's open agent framework designed to solve GitHub issues autonomously. - **[OpenHands (formerly OpenDevin)](https://github.com/All-Hands-AI/OpenHands)** `MIT` `46k Stars` — Open-source community project building autonomous software development agents. - **[JetBrains Junie](https://www.jetbrains.com/junie/)** `Commercial` — Embedded agent for IntelliJ, PyCharm, and WebStorm with native MCP and codebase grounding. - **[Amazon Q Developer Agent](https://aws.amazon.com/q/developer/)** `Commercial` — AWS's enterprise engineering agent capable of executing multi-repository feature implementations, dependency upgrades, and cloud security remediations in isolated sandboxes. - **[Plandex](https://github.com/plandex-ai/plandex)** `AGPL-3.0` — Open-source, terminal-native AI coding engine designed for complex, multi-file code tasks with tree-based planning and long-horizon execution. - **[AutoCodeRover](https://github.com/nus-apr/auto-code-rover)** `Apache 2.0` — Automated software engineering agent using AST and program structure analysis to navigate codebases and resolve complex GitHub issues. - **[Agent QA](https://github.com/vostride/agent-qa)** `FSL-1.1-ALv2` — Self-improving QA agent for natural-language web and mobile regression tests with persistent test memory, a dashboard, CLI, MCP server, and agent skills. - **[Qodo (formerly CodiumAI)](https://www.qodo.ai/)** `Commercial` `Agentic Testing` — Agentic code analysis platform offering automated test generation, PR review agents, and context-aware integrity verification. - **[YYLO](https://github.com/yylo-dev/yylo)** `MIT` — Open-source command-line orchestrator for coding agents that runs each task in a dedicated branch/worktree behind typed task, validation, merge, and release-readiness boundaries, with a risk-based merge queue and receipt-backed repository changes. ### 6.2. DevOps, SRE & Cloud Infrastructure - **[Kubiya](https://www.kubiya.ai)** `Commercial` — Conversational DevOps and infrastructure automation agent executing operations across Kubernetes, Terraform, and cloud platforms. - **[Rootly AI SRE Agent](https://rootly.com)** `Commercial` — Autonomous SRE and incident response agent automating incident triage, status page communications, and post-mortem generation. - **[K8sGPT](https://k8sgpt.ai/)** `Apache 2.0` `K8s Remediation` — Kubernetes SRE agent that scans clusters, diagnoses complex YAML/runtime errors in plain English, and provides automated remediation actions. - **[PagerDuty Advance Agents](https://www.pagerduty.com/platform/generative-ai/)** `Commercial` — Enterprise incident triage agent automating responder mobilization, runbook diagnostics, and incident timeline summaries. - **[Shoreline AI](https://shoreline.io/)** `Commercial` — Cloud reliability agent that auto-detects operational anomalies and triggers autonomous remediation workflows across production fleets. - **[Dynatrace Davis CoPilot](https://www.dynatrace.com/)** `Commercial` — Observability agent leveraging deterministic topology graphs to automate root-cause analysis and cloud remediation. ### 6.3. Deep Research & Intelligence - **[Perplexity Deep Research](https://www.perplexity.ai/)** `Proprietary` `Multi-Hop Search` — Autonomous multi-step research agent executing dozens of web queries, citations, and analytical synthesis to produce exhaustive reports. - **[OpenAI Deep Research](https://openai.com/)** `Proprietary` `Long-Form Papers` — Deep research agent that browses hundreds of web sources asynchronously and generates long-form research papers. - **[GPT Researcher](https://github.com/assafelovic/gpt-researcher)** `Apache 2.0` `20k Stars` — Autonomous open-source agent that executes parallel web scraping, aggregates 20+ sources, and compiles comprehensive research papers with citations. - **[Stanford STORM](https://github.com/stanford-oval/storm)** `MIT` — Synthesis of Topic Outlines through Grounded LLM Multi-Agent Interaction for generating structured, Wikipedia-style deep research reports. - **[Hebbia Matrix](https://www.hebbia.ai/)** `Commercial` — Enterprise research agent analyzing thousands of financial, legal, and unstructured documents simultaneously in a spreadsheet interface. - **[Elicit Research Agent](https://elicit.com/)** `Commercial` — Automated literature review and scientific synthesis agent analyzing over 200 million academic papers. - **[Consensus](https://consensus.app/)** `Commercial` — Research agent that searches academic literature, synthesizes consensus across clinical/scientific studies, and extracts evidence. ### 6.4. Business, Finance & HR - **[Statsnet](https://github.com/usenetstate/statsnet-mcp)** — Background check any company in the world: registration, executives, courts and finances. Endpoint: `https://statsnet.co/mcp` · Registry: `io.github.usenetstate/statsnet` - **[Salesforce Agentforce](https://salesforce.com/agentforce)** `Commercial` `Enterprise Swarm` — Autonomous enterprise agent suite across CRM, sales pipeline, service triage, and marketing workflows. - **[Claude Cowork](https://claude.ai)** `Anthropic` `Knowledge Ops` — Desktop agentic platform for knowledge work — reconciling spend, building trackers from contract folders, and assembling decks from transcripts across connected apps. - **[Workday Sana](https://workday.com)** `Commercial` — Enterprise AI agent suite automating HR, payroll, procurement, and financial operations. - **[SAP Joule Studio 2.0](https://sap.com)** `Commercial` — Enterprise agent layer for supply chain, HCM, and ERP business processes across SAP environments. - **[Microsoft Agent 365](https://microsoft.com)** `Commercial` — Control plane for building, deploying, and governing business agents inside Microsoft 365 environments. - **[Glean](https://glean.com)** `Commercial` — Enterprise search and autonomous agent engine connected across Slack, Google Drive, Jira, and Notion. - **[FinRobot](https://github.com/AI4Finance-Foundation/FinRobot)** `MIT` `Quantitative Agent` — Open-source financial AI agent platform designed for quantitative analysis, equity research, algorithmic trading, and SEC filing synthesis. ### 6.5. Customer Experience & Support - **[Intercom Fin](https://www.intercom.com/fin)** `Commercial` `Auto-Resolution` — AI customer support agent resolving user queries without human intervention across chat and email. - **[Zendesk AI Agents](https://www.zendesk.com/service/ai/ai-agents/)** `Commercial` — Customer service agent suite with automated resolution workflows across omnichannel tickets. - **[Sierra](https://sierra.ai)** `Commercial` `Brand Supervisor` — Enterprise CX agent platform utilizing supervisor agent patterns to maintain brand alignment and safe execution. - **[Ada CX Platform](https://www.ada.cx/)** `Commercial` — Enterprise automated customer service platform resolving multi-turn requests across voice, web, and messaging apps. - **[Decagon](https://decagon.ai)** `Commercial` — Voice, chat, and email CX agents built with continuous testing and evaluation loops. - **[Cognigy.AI](https://www.cognigy.com/)** `Commercial` — Enterprise conversational AI platform orchestrating contact center agent swarms and telephony pipelines. - **[Maven AGI](https://www.mavenagi.com)** `Commercial` — Multi-channel enterprise support agent engine with real-time knowledge synthesis. ### 6.6. Work Management & Enterprise Ops - **[ServiceNow AI Agent Studio](https://www.servicenow.com)** `Commercial` — Enterprise workflow agent builder for IT service management (ITSM), HR delivery, and ops automation. - **[Atlassian Rovo Agents](https://www.atlassian.com/rovo)** `Commercial` — Specialized AI teammates integrated into Jira, Confluence, and enterprise work graphs. - **[Asana AI Teammates](https://asana.com/product/ai/ai-teammates)** `Commercial` — Collaborative agents that track goals, follow up on deliverables, and assign work inside Asana. - **[ClickUp Brain Agents](https://clickup.com/brain)** `Commercial` — Context-aware workspace agents for project planning, doc writing, and task management. - **[Notion AI Agents](https://www.notion.so/product/ai)** `Commercial` — Integrated workspace agent that plans projects, searches cross-document context, and synthesizes action items. - **[Monday.com AI Automations](https://monday.com/ai)** `Commercial` — Adaptive work management agents handling project tracking, dependency management, and automated status reporting. ### 6.7. Science, BioTech & Materials - **[CAS Newton](https://cas.org)** `Commercial` — Deep scientific discovery agent grounded in 150+ years of chemical and biological research data. - **[NVIDIA BioNeMo](https://nvidia.com/bionemo)** `Research` — Generative AI and agent framework for biomolecular discovery, protein structure prediction, and drug target validation. - **[AlphaFold 3 Agent](https://deepmind.google)** `Research` `Structural Biology` — DeepMind's structural biology system modeling joint structures of proteins, DNA, RNA, and ligands. - **[EvolutionaryScale ESM3](https://www.evolutionaryscale.ai/)** `Open Weights` `Generative Biology` — Frontier biological foundation model generating novel proteins, catalytic sequences, and functional enzymes from first principles. - **[Evo (Arc Institute)](https://github.com/evo-design/evo)** `Apache 2.0` `CRISPR & DNA` — Biological foundation model and generative agent capable of designing synthetic DNA, CRISPR systems, and protein assemblies. - **[ChemCrow](https://github.com/ur-whitelab/chemcrow-public)** `MIT` `Autonomous Chemistry` — Chemistry agent integrating 18 expert-designed computational chemistry tools for organic synthesis and reaction planning. - **[Abridge](https://www.abridge.com)** `Healthcare` — Clinical documentation agent structuring doctor-patient encounters into clinical notes and EHR updates. - **[Ambience Healthcare](https://www.ambiencehealthcare.com)** `Healthcare` — Operating system for AI clinical documentation, coding, and compliance workflows. - **[Hippocratic AI](https://www.hippocraticai.com)** `Healthcare` — Safety-focused non-diagnostic healthcare agents for patient follow-up, scheduling, and health education. - **[BioMistral](https://github.com/BioMistral/BioMistral)** `Apache 2.0` `Clinical LLM` — Medical-domain open LLM agent suite tailored for healthcare reasoning, clinical Q&A, and EHR integration. ### 6.8. Data Analytics & Mathematics - **[Julius AI](https://julius.ai)** `Freemium` — Interactive data science agent running Python/R scripts to analyze spreadsheets, generate charts, and build predictive models. - **[Databricks Genie](https://docs.databricks.com/en/genie/)** `Commercial` — Governed enterprise data agent grounded in Unity Catalog data lakes and workspace intelligence. - **[Snowflake Cortex Agents](https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents)** `Commercial` — Autonomous analytics agents querying structured tables and unstructured documents inside Snowflake. - **[Microsoft Fabric Data Agents](https://learn.microsoft.com/en-us/fabric/data-science/concept-data-agent)** `Commercial` — Conversational data agents for OneLake, Power BI, and Microsoft Graph. - **[Vanna.ai](https://github.com/vanna-ai/vanna)** `MIT` `14k Stars` — Open-source RAG-based SQL generation agent that trains on database schema, metadata, and DDL to produce precise analytical queries. - **[PandasAI](https://github.com/gventuri/pandas-ai)** `MIT` `18k Stars` — Conversational data analysis agent turning natural language into executable Pandas and SQL analytics workflows. ### 6.9. Virtual Worlds & Gaming - **[Altera](https://altera.al)** `Commercial` — Autonomous digital humans operating in virtual sandbox worlds (e.g. Minecraft), demonstrating emergent social behaviors and economies. - **[Oasis Gen-World](https://oasis.ai)** `Research` — Real-time world generation engine creating infinite interactive game environments on the fly. - **[Inworld AI Engine 3.0](https://inworld.ai)** `Commercial` — Dynamic NPC brain framework providing persistent memory, emotion models, and adaptive narrative execution. - **[Voyagers / MineDojo](https://github.com/MineDojo/Voyager)** `MIT` — Lifelong learning agent in Minecraft driven by LLMs, continuous skill libraries, and automatic curricula. ### 6.10. Physical AI (Robotics) - **[NVIDIA Cosmos 3](https://developer.nvidia.com/cosmos)** `Open Weights` `World Foundation` — Foundation world model for physical AI, robotics simulation, and synthetic dataset generation. - **[NVIDIA Isaac GR00T N1.7](https://developer.nvidia.com/isaac/groot)** `Open Weights` `VLA Foundation` — Vision-Language-Action (VLA) foundation reasoning model built specifically for humanoid robots. - **[Boston Dynamics Atlas](https://bostondynamics.com/atlas)** `Commercial` `Electric Humanoid` — Fully electric humanoid robot powered by VLA physical reasoning stacks for industrial manufacturing. - **[Figure 02](https://figure.ai)** `Commercial` — Commercial humanoid worker agent deployed in automotive manufacturing lines (BMW/Mercedes). - **[Unitree G1 / H1](https://www.unitree.com/g1)** `Commercial` — High-agility mass-production humanoid robot utilizing multimodal reinforcement learning and visual terrain negotiation. - **[1X Neo](https://www.1x.tech/)** `Commercial` — Bipedal android designed for consumer and enterprise physical assistance powered by embodied AI foundation models. - **[LG CLOi](https://lg.com/global/robot)** `Commercial` — Smart home humanoid robot using edge AI reasoning to execute household physical chores. ### 6.11. Law, Legal & Compliance - **[Harvey AI](https://harvey.ai)** `Commercial` — Legal reasoning agent platform for law firms and corporate legal teams, automating contract analysis, redlining, and research. - **[Robin AI](https://robinai.co.uk)** `Commercial` — Contract negotiation agent automatically evaluating standard agreement terms and drafting revisions. - **[LexisNexis Lexis+ AI Agent](https://lexisnexis.com)** `Commercial` — Legal research agent generating verified briefs grounded in primary legal authorities. - **[Casetext CoCounsel](https://casetext.com/cocounsel)** `Commercial` — Legal AI assistant acquired by Thomson Reuters, performing document review, legal research memos, and deposition preparation. ### 6.12. Creative Media & Design - **[Figma AI / Genius Agent](https://figma.com)** `Commercial` — Collaborative UI/UX design agent generating, editing, and componentizing dynamic design systems inside Figma. - **[Runway Gen-3 / Sora Media Agent](https://runwayml.com)** `Commercial` — Multimodal creative video agent automating multi-shot storyboarding, character consistency, and spatial editing. - **[Uizard AI / Galatea Agent](https://uizard.io)** `Freemium` — Autonomous UI design agent transforming wireframe sketches and natural language prompts into interactive app prototypes. - **[Midjourney Agentic Canvas](https://www.midjourney.com)** `Commercial` — Visual generative agent allowing parametric inpainting, regional reframing, and multi-asset art direction. ### 6.13. Education & Adaptive Learning - **[Khanmigo](https://www.khanacademy.org/khanmigo)** `Freemium` `Socratic Agent` — Khan Academy's interactive AI tutoring agent providing Socratic step-by-step guidance across STEM, computer science, and humanities. - **[Synthesia Persona](https://www.synthesia.io)** `Commercial` — Interactive avatar tutoring agents for automated corporate training, language learning, and adaptive curriculum delivery. - **[Duolingo Max Agent](https://www.duolingo.com/)** `Freemium` — Conversational roleplay and explanation agent providing adaptive, personalized language mastery. ### 6.14. Cybersecurity & Threat Intelligence - **[Google Big Sleep / Project Naptime](https://googleprojectzero.blogspot.com/)** `Research` `Zero-Day Agent` — Google Project Zero's autonomous agent framework that successfully discovered, triaged, and verified real-world in-the-wild zero-day vulnerabilities in production codebases. - **[Caido Agent](https://caido.io/)** `Commercial` `Offensive AppSec` — Autonomous web penetration testing and API security agent that actively explores attack surfaces and discovers injection flaws. - **[HexStrike](https://github.com/hexstrike/hexstrike)** `Open Source` `Automated Red Team` — Open-source autonomous adversary simulation agent executing MITRE ATT&CK tactical red-teaming campaigns. - **[CrowdStrike Charlotte AI](https://crowdstrike.com)** `Commercial` `SOC Defense` — Real-time autonomous SOC agent inside Falcon platform automating cross-endpoint threat hunting and remediation. - **[Palo Alto Prisma AIRS 3.0](https://paloaltonetworks.com/prisma/airs)** `Commercial` `Agent Gateway` — Autonomous AI runtime security platform providing inline prompt firewalling, tool isolation, and OWASP ASI Top 10 mitigation. - **[VulnCheck Agent](https://vulncheck.com/)** `Commercial` — Autonomous exploit intelligence and vulnerability tracking agent correlating real-time exploit chains across open-source dependencies. ### 6.15. Supply Chain & Autonomous Logistics - **[Project44 Movement AI](https://www.project44.com/)** `Commercial` — Autonomous multimodal supply chain agent predicting freight disruptions, rerouting shipments, and mitigating port delays. - **[Flexport Autonomous Freight Agent](https://www.flexport.com/)** `Commercial` — Logistics agent automating customs filing, ocean/air quote negotiation, and shipment milestone tracking. - **[Manhattan Associates Supply Agent](https://www.manh.com/)** `Commercial` — Autonomous warehouse management agent optimizing slotting, inventory replenishment, and robotics dispatch. - **[Kinaxis Maestro](https://www.kinaxis.com/)** `Commercial` — Supply chain orchestration engine providing real-time multi-tier scenario modeling and automated inventory rebalancing. ### 6.16. Sales, Go-To-Market (GTM) & Revenue Swarms - **[11x (Alice & Jordan)](https://11x.ai/)** `Commercial` `Autonomous SDR` — Autonomous digital sales development reps that research prospects, craft personalized multi-channel outreach, handle objections, and schedule meetings. - **[Artisan (Ava AI BDR)](https://artisan.co/)** `Commercial` `Outbound Agent` — Autonomous business development agent conducting lead enrichment, warm email deliverability orchestration, and intent-driven prospecting. - **[Clay / Claygent](https://www.clay.com/)** `Commercial` `GTM Research Agent` — Autonomous web-scraping and data enrichment agent crawling corporate websites, LinkedIn, and GitHub to build verified lead graphs. - **[Regie.ai](https://www.regie.ai/)** `Commercial` — Agentic prospecting platform that monitors intent signals and dynamically generates tailored sales campaigns for revenue teams. - **[Qualified Piper](https://www.qualified.com/)** `Commercial` — Inbound website conversational agent that qualifies enterprise buyers in real-time and routes high-intent pipeline directly into sales calendars. ### 6.17. Defense, Aerospace & National Security - **[Palantir AIP Agents](https://www.palantir.com/platforms/aip/)** `Commercial` `Mission Intelligence` — Governed enterprise agent orchestration platform connecting operational ontologies, mission data, and defense action workflows with strict access controls. - **[Anduril Lattice Agentic OS](https://www.anduril.com/lattice/)** `Commercial` `Autonomous Command` — Real-time command-and-control software coordinating autonomous sensor grids, unmanned aerial systems (UAS), and tactical defense assets. - **[Scale AI Donovan](https://scale.com/donovan)** `Commercial` `Defense LLM` — Decision-support agent platform built for national security and defense personnel to synthesize satellite imagery, signals, and multi-source intel. - **[Primer AI](https://primer.ai/)** `Commercial` — Natural language intelligence agent analyzing open-source intelligence (OSINT) and classified data feeds to deliver real-time situational threat briefings. ### 6.18. Energy, Utilities & Climate Tech - **[AutoGrid / Uplight](https://www.auto-grid.com/)** `Commercial` `Virtual Power Plant` — Autonomous energy flexibility agent forecasting grid demand, balancing distributed energy resources, and bidding storage assets in wholesale power markets. - **[Gridmatic](https://www.gridmatic.com/)** `Commercial` `AI Power Trading` — Algorithm-driven autonomous energy trading agent operating on renewable generation and battery storage dispatch. - **[Pachama](https://pachama.com/)** `Commercial` `Satellite Carbon Verification` — Geospatial agent utilizing satellite LIDAR and computer vision to verify carbon offset projects and detect deforestation in near real-time. - **[LevelTen Energy AI](https://www.leveltenenergy.com/)** `Commercial` — Renewable energy procurement agent analyzing power purchase agreements (PPAs) and risk-pricing clean energy transactions. ### 6.19. Insurance & Risk Underwriting - **[Sixfold](https://sixfold.am/)** `Commercial` `Underwriting Co-Pilot` — Purpose-built generative agent for commercial insurance underwriters, synthesizing regulatory filings, risk portfolios, and capacity limits. - **[Shift Technology](https://www.shift-technology.com/)** `Commercial` `Fraud Adjudication` — Autonomous insurance fraud detection and claims decisioning agent analyzing behavioral telemetry and healthcare billing codes. - **[Hyperscience](https://www.hyperscience.com/)** `Commercial` — Hyperautomation agent parsing complex handwriting, unstructured policy documents, and claims submissions with automated human-in-the-loop fallback. - **[ClaimSorted](https://www.claimsorted.com/)** `Commercial` — End-to-end autonomous medical claims adjudication agent reviewing health receipts against dynamic insurance policy rule sets. [Back to Top](#nexum--the-agentic-universe) --- ## 07. Agent Evaluation, Tracing & Observability *Measuring accuracy, latency, execution trajectories, unit assertions, and cost.* ### 7.1. Benchmarks & Leaderboards - **[SWE-bench / SWE-bench Verified / SWE-bench Pro](https://www.swebench.com/)** `Standard` — The industry benchmark evaluating software engineering agents on real GitHub issues from popular open-source repositories. - **[Terminal-Bench 2.1](https://www.tbench.ai/)** `Terminal Benchmark` — Fast-rising standard for evaluating CLI/terminal coding agents (Codex CLI, Claude Code, Cursor, Antigravity) via the open Harbor evaluation framework. - **[GAIA (General AI Assistants)](https://huggingface.co/spaces/gaia-benchmark/leaderboard)** `Multimodal Benchmark` — Benchmark testing multimodal AI agents on complex, multi-step real-world tasks requiring browsing, calculation, and tool use. - **[OSWorld](https://os-world.github.io/)** `OS Benchmark` — Operating system interaction benchmark assessing agents' visual screen perception and mouse/keyboard execution on Windows, Ubuntu, and macOS. - **[WebArena & Tau-Bench](https://webarena.dev/)** `Web Benchmark` — Web-based environment benchmarks evaluating autonomous web navigation, e-commerce transactions, and customer service workflows. - **[Humanity's Last Exam (HLE)](https://cais.community/hle/)** `Frontier Benchmark` — Frontier multimodal benchmark designed to evaluate advanced agentic reasoning across academic and technical disciplines. ### 7.2. Datasets, Golden Traces & Trajectory Corpora *Curated training and evaluation datasets containing multi-step agent interaction traces.* - **[SWE-bench Benchmark Dataset](https://github.com/princeton-nlp/SWE-bench)** `Princeton` — 2,294 real-world software engineering problem instances extracted from 12 popular Python repositories with unit test assertions. - **[ToolBench / ToolBench 2](https://github.com/OpenBMB/ToolBench)** `OpenBMB` — Massive dataset containing 16,000+ real-world REST APIs from RapidAPI, featuring 120,000+ multi-step agent tool calling execution trajectories. - **[Mind2Web Dataset](https://github.com/OSU-NLP-Group/Mind2Web)** `OSU NLP` — Comprehensive dataset for web agents spanning 2,000+ open-ended tasks across 137 websites with ground-truth DOM action sequences. - **[WebArena Environment & Traces](https://github.com/web-arena-x/webarena)** `CMU` — Standalone web environment simulation dataset spanning e-commerce, forums, and GitLab tasks for evaluating autonomous agents. - **[AgentDrive Trajectories](https://github.com/autonomousvision/agent-drive)** `Research` — Trajectory dataset for autonomous decision-making and spatial reasoning in complex dynamic environments. ### 7.3. Tracing, Telemetry & Evaluation Suites - **[LangSmith](https://www.langchain.com/langsmith)** `Commercial` `Tracing Standard` — Industry-standard developer platform for tracing, testing, evaluating, and monitoring LLM applications and agentic graphs. - **[Langfuse](https://langfuse.com)** `MIT` `9k Stars` — Open-source observability and evaluation platform for AI agents, featuring detailed trace visualizations, cost tracking, and prompt management. - **[Arize Phoenix](https://phoenix.arize.com)** `Open Source` `5k Stars` — AI observability platform built on OpenTelemetry for tracing agent execution loops, evaluation metrics, and tool execution call stacks. - **[AgentOps](https://www.agentops.ai)** `Freemium` `Session Replay` — Purpose-built observability platform for multi-agent systems, tracking agent session replays, LLM costs, tool failures, and recursion loops. - **[DeepEval / DeepTeam](https://github.com/confident-ai/deepeval)** `Apache 2.0` `7k Stars` — Open-source LLM & agent unit testing and red-teaming framework, featuring OWASP ASI Top 10 test coverage. - **[Promptfoo](https://github.com/promptfoo/promptfoo)** `MIT` `7k Stars` — CLI & library for evaluating agent prompts, tool calls, agentic security vulnerabilities, and automated red-teaming. - **[Ragas](https://github.com/explodinggradients/ragas)** `Apache 2.0` `10k Stars` — Framework for evaluating Retrieval Augmented Generation (RAG) pipelines and multi-turn conversational agents with automated assertion metrics. - **[TruLens](https://github.com/truera/trulens)** `Apache 2.0` — Evaluation and instrumentation tool suite providing feedback functions to benchmark agent grounding, relevance, and toxicity. - **[Braintrust](https://www.braintrust.dev/)** `Commercial` — Enterprise evaluation stack for automated agent testing, prompt playground experiments, and CI/CD quality gates. - **[Comet Opik](https://www.comet.com/site/products/opik/)** `Open Source` — End-to-end LLM evaluation and agent tracing platform supporting custom metrics, automated evaluation, and real-time monitoring. - **[Traceloop (OpenLLMetry)](https://www.traceloop.com/)** `Open Source` — OpenTelemetry-native telemetry standard for monitoring agent traces across OpenTelemetry collectors (Datadog, Dynatrace, New Relic). - **[YYLO Benchmark](https://github.com/yylo-dev/yylo-benchmark)** `MIT` — Flexible, isolated evaluation layer for task prompts and Workflow Runner YAML that normalizes every case into one attempt contract, runs each candidate in a private fresh-repository workspace, and evaluates retained evidence with ordered deterministic and/or LLM profiles, retaining hash-verified terminals and immutable records. [Back to Top](#nexum--the-agentic-universe) --- ## 08. Guardians: Security & Governance *Ensuring the autonomous world stays safe, aligned, and compliant.* > [!NOTE] > **The Reference Security Standard:** The [OWASP Top 10 for Agentic Applications (ASI Top 10)](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/), published with 100+ contributing organizations, is the industry baseline. It enforces **Least-Agency** (granting only necessary autonomy) and **Strong Observability** (mandatory audit logging of goal state and tool paths). ### 8.1. Security Gateways & Guardrails - **[Witness AI](https://witness.ai/)** `Commercial` — Enterprise security control plane monitoring and intercepting agentic intent, preventing unauthorized system actions. - **[Cisco Splunk AI SOC Agents](https://cisco.com/splunk)** `Commercial` — Specialized security operations agents automating alert triage, incident investigation, and threat mitigation with OWASP ASI Top 10 coverage. - **[CrowdStrike Charlotte AI](https://crowdstrike.com)** `Commercial` — Agentic SOC platform inside Falcon SIEM automating threat hunting, triage, and mitigation routines. - **[Palo Alto Prisma AIRS 3.0](https://paloaltonetworks.com/prisma/airs)** `Commercial` — Security suite covering full agent lifecycles: pre-deployment scanning, inline prompt protection, and red-teaming, explicitly mapped to the OWASP ASI Top 10. - **[Wiz Security Agents](https://wiz.io)** `Commercial` — Cloud security guardrails designed specifically for AWS AgentCore, Gemini Enterprise Agent Platform, and Salesforce Agentforce. - **[Garak (LLM Vulnerability Scanner)](https://github.com/leondz/garak)** `Apache 2.0` `4k Stars` — Open-source vulnerability scanner and automated red-teaming probe discovering hallucinations, jailbreaks, and injection flaws in AI agents. - **[HiddenLayer AISec](https://hiddenlayer.com/)** `Commercial` — Enterprise security platform protecting AI systems from adversarial attacks, model theft, prompt injections, and rogue agent behavior. - **[Llama Guard 3](https://llama.meta.com)** `Open Weights` — Meta's safety classification model for input and output agent moderation across text, code, and tool execution. - **[Microsoft Agent Governance Toolkit](https://github.com/microsoft/agent-governance-toolkit)** `MIT` — Open-source runtime toolkit enforcing safety constraints, identity boundaries, and OWASP compliance on AI agents. - **[Lakera Guard](https://www.lakera.ai)** `Commercial` — High-speed enterprise guardrail intercepting prompt injections, jailbreaks, and sensitive data leakage in real-time. - **[NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails)** `Apache 2.0` — NVIDIA's open-source toolkit for adding programmable safety, security, and topical rails to agentic LLMs. - **[Guardrails AI](https://www.guardrailsai.com/)** `Apache 2.0` — Validation framework ensuring agent outputs conform to strict structural, safety, and semantic constraints. - **[NeuralTrust](https://neuraltrust.ai)** `Commercial` — Agentic-security platform built directly around the OWASP ASI Top 10 (Goal Hijack, Tool Misuse, Privilege Abuse, Memory Poisoning). ### 8.2. OWASP Top 10 for Agentic Applications (ASI Top 10) | Risk ID | Vulnerability Classification | Primary Threat Vector | Architectural Mitigation | | :--- | :--- | :--- | :--- | | **ASI01** | **Agent Goal Hijack** | Prompt injection altering the agent's core goal | Deterministic guardrails & plan-validation gateways | | **ASI02** | **Tool Misuse & Unsafe Execution** | Agent executing destructive OS or shell commands | Strict sandbox isolation & Human-in-the-Loop (HITL) | | **ASI03** | **Privilege & Scope Escalation** | Agent accessing resources beyond authorized boundaries | Least-Agency enforcement & scoped token minting | | **ASI04** | **Memory & Context Poisoning** | Malicious injection into persistent episodic vector storage | Memory sanitization & cryptographic provenance checks | | **ASI05** | **Cascading Swarm Failures** | Error amplification across cyclic multi-agent graphs | Circuit breakers & max-depth recursion limits | | **ASI06** | **Excessive Agency & Hallucinated Actions** | Agent acting without cryptographic confirmation | Explicit user mandate signatures (e.g. AP2) | | **ASI07** | **Insecure Tool Outputs & Handoffs** | Downstream injection from untrusted API/scraping output | Structured schema validation before context ingestion | | **ASI08** | **Supply Chain & Skill Poisoning** | Compromised third-party MCP servers or agent plugins | MCP server attestation & checksum verification | | **ASI09** | **Data Exfiltration via Autonomous Side-Channels**| Agent transmitting secrets via unauthorized web endpoints | Strict egress firewalling & API gateway allowlists | | **ASI10** | **Rogue / Unkillable Agent Loops** | Infinite recursion consuming unconstrained API budget | Hard compute/dollar budget caps & mandatory heartbeat monitors | ### 8.3. Automated Red-Teaming & Adversarial Probing - **[Microsoft PyRIT (Python Risk Identification Tool)](https://github.com/Azure/PyRIT)** `MIT` — Open-source automation framework used by AI red teams to assess agent vulnerabilities against adversarial prompts, multi-turn jailbreaks, and sensitive data leakage. - **[DeepEval RedTeam Suite](https://github.com/confident-ai/deepeval)** `Apache 2.0` — Synthetic adversarial attack generator probing agent tool calls against OWASP ASI Top 10 vectors. - **[Promptfoo RedTeam CLI](https://github.com/promptfoo/promptfoo)** `MIT` — Automated fuzzing tool generating thousands of boundary attacks, indirect prompt injections, and privilege-escalation probes. [Back to Top](#nexum--the-agentic-universe) --- ## 09. Agentic Commerce & Payments Protocols *Production rails enabling autonomous agents to execute financial transactions.* > [!TIP] > **The Transacting Agent Economy:** Analysts project agentic commerce will influence $3–5 trillion in transactions by 2030. Modern architectures separate user authorization (mandates), agent-to-agent negotiation, and settlement layers. ### 9.1. Commerce Protocols & Wallets - **[Agent Payments Protocol (AP2)](https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol)** `Google` `Mandates Standard` — An authorization and trust framework using cryptographically signed "mandates" to prove human user authorization for agent purchases. Backed by 60+ partners including Amex, Coinbase, Mastercard, PayPal, and Salesforce. - **[x402 Protocol](https://github.com/coinbase/x402)** `Coinbase / Linux Foundation` `HTTP 402` — Open protocol reviving HTTP 402 for instant stablecoin (USDC) micropayments between agents and APIs without accounts or API keys. Co-governed by the Linux Foundation and Cloudflare. - **[Nano (XNO) via x402](https://github.com/x402nano/exact)** `Nano` `Feeless L1 Settlement` — Nano is an instant, feeless, green layer-1 that settles fixed-amount XNO micropayments over the x402 protocol (Client / Resource Server / Facilitator) with zero fees on either side. - **[A2A x402 Extension](https://developers.google.com)** `Google` — Bridges Google's A2A protocol to x402 settlement, enabling agents to discover services, authorize via AP2, and settle via x402 in a single flow. - **[Agentic Commerce Protocol (ACP)](https://openai.com)** `OpenAI` `Instant Checkout` — Powers ChatGPT Instant Checkout as an open protocol supported by Stripe, Shopify, Salesforce, and PayPal. - **[Multi-Party Payments (MPP)](https://stripe.com)** `Consortium` — Settlement rail supporting multi-party transaction splits across Stripe, Visa, Mastercard, Anthropic, OpenAI, and Shopify. - **[Amazon Bedrock AgentCore Payments](https://aws.amazon.com/bedrock/agentcore/)** `AWS` — Managed AWS capability enabling Bedrock agents to negotiate x402 payments, authenticate wallets, and enforce spending caps. - **[Coinbase Agent Wallet](https://coinbase.com)** `Infrastructure` — On-chain wallet infrastructure empowering agents to hold crypto, pay for compute, and settle smart contracts. - **[Crossmint](https://www.crossmint.com)** `Commercial` — Enterprise agent wallet platform unified across x402, AP2 mandates, and conversational checkout with programmable guardrails. - **[Skyfire](https://skyfire.xyz)** `Financial Infrastructure` — Banking and payment network enabling AI agents to hold balances and pay merchant endpoints. - **[AffixIO](https://www.npmjs.com/package/affixio)** `npm` `Agentic Pay` - Agentic Pay Kit. Install with `npm i affixio`. Host-side signed ACTION attestation before pay (x402BeforePay / Agentic Pay / KYA; not person KYC). ([GitHub](https://github.com/AffixIO/SDK)) ([docs](https://www.affix-io.com/agent-trust/)) ### 9.2. Agentic Payments Protocol Matrix | Protocol | Governance / Backing | Architectural Layer | Primary Asset / Currency | Key Innovation | | :--- | :--- | :--- | :--- | :--- | | **AP2** | Google, Amex, PayPal, Mastercard | Authorization & Intent | Multi-Currency / Fiat / Tokens | Cryptographically signed user mandates for agent purchases | | **x402** | Coinbase, Linux Foundation, Cloudflare | HTTP Settlement Layer | USDC / Stablecoins | Revives HTTP 402 for sub-cent API micropayments with zero API keys | | **Nano (XNO)** | Nano Foundation; x402nano | Feeless L1 Settlement | XNO | Instant, feeless, green layer-1 settlement for fixed-amount x402 micropayments | | **ACP** | OpenAI, Stripe, Shopify, Salesforce | Commercial Checkout | Fiat & Card Rails | Instant checkout protocol for conversational shopping agents | | **MPP** | Stripe, Visa, Anthropic, OpenAI | Multi-Party Settlement | Multi-Asset Rails | Automated revenue splits across multi-agent collaborative swarms | > [!WARNING] > **Security Precaution:** Agentic payment rails require strict governance. Always enforce the principle of Least-Agency and configure strict spending caps on agent wallet credentials. [Back to Top](#nexum--the-agentic-universe) --- ## 10. Agent Operations & Economy *Workflow tools, orchestration platforms, and sub-cent request metering.* ### 10.1. Workflow & Low-Code - **[n8n](https://n8n.io)** `Fair-code` `102k Stars` — Leading self-hosted workflow automation platform with native MCP tools, A2A protocols, and visual multi-agent builders. - **[Dify](https://dify.ai)** `MIT` `108k+ Stars` — Open-source LLM app development platform supporting multi-agent workflow graphs, RAG pipelines, and model routing. - **[Langflow](https://langflow.org)** `MIT` `45k Stars` — Visual drag-and-drop agent framework with native LangChain, MCP, and multi-agent graph capabilities. - **[Flowise](https://github.com/FlowiseAI/Flowise)** `MIT` `34k Stars` — Drag-and-drop UI to build customized LLM flows, conversational agents, and tool-augmented multi-agent swarms. - **[AutoKitteh](https://github.com/autokitteh/autokitteh)** `Apache 2.0` — Developer platform for orchestrating serverless, event-driven agent workflows and interactive bots across external APIs. - **[Lindy](https://lindy.ai)** `Commercial` — No-code agent builder automating personal and business task workflows across email, calendar, CRM, and communication apps. - **[Bolt.new / v0](https://bolt.new)** `Commercial` — Full-stack web application generation environments driven by autonomous code agents. - **[OpenWebUI Agent Hub](https://github.com/open-webui/open-webui)** `MIT` `90k+ Stars` — Self-hosted web interface enabling multi-agent orchestration, web browsing tools, RAG memory pipelines, and model-agnostic agent delegation. ### 10.2. Metering & Billing - **[Stripe Agent Billing](https://stripe.com/agents)** `Commercial` `Micro-Billing` — API infrastructure enabling AI agents to bill for compute usage, manage subscriptions, and process micro-transactions; interoperable with ACP, x402, and MPP. - **[Helicone](https://helicone.ai)** `Apache 2.0` `Cost Management` — LLM observability and cost management platform providing usage throttling, API rate-limiting, and cost allocation per agent session. - **[Nevermined](https://nevermined.ai)** `Commercial` — Purpose-built billing platform for sub-cent, per-request agent metering built on top of x402-style settlement. [Back to Top](#nexum--the-agentic-universe) --- ## 11. Seminal Research Papers & Literature *Foundational academic papers establishing the theoretical and empirical bedrock of autonomous agents.* | Year | Paper Title | Authors / Institution | Core Breakthrough | Link | | :--- | :--- | :--- | :--- | :--- | | **2022** | *ReAct: Synergizing Reasoning and Acting in Language Models* | Yao et al. (Princeton & Google Brain) | Interleaved thought and action traces enabling tool use | [arXiv:2210.03629](https://arxiv.org/abs/2210.03629) | | **2023** | *Generative Agents: Interactive Simulacra of Human Behavior* | Park et al. (Stanford University) | Multi-agent sandbox with episodic memory and reflection | [arXiv:2304.03442](https://arxiv.org/abs/2304.03442) | | **2023** | *Reflexion: Language Agents with Verbal Reinforcement Learning* | Shinn et al. (Northeastern & MIT) | Self-evaluative verbal memory for error self-correction | [arXiv:2303.11366](https://arxiv.org/abs/2303.11366) | | **2023** | *Toolformer: Language Models Can Teach Themselves to Use Tools* | Schick et al. (Meta AI Research) | Self-supervised learning of API and tool invocation | [arXiv:2302.04761](https://arxiv.org/abs/2302.04761) | | **2024** | *SWE-bench: Can Language Models Resolve Real-World GitHub Issues?* | Jimenez et al. (Princeton University) | Industry-standard benchmark for software engineering agents | [arXiv:2310.06770](https://arxiv.org/abs/2310.06770) | | **2024** | *OSWorld: Benchmarking Multimodal Agents on OS Interaction* | Xie et al. (HKU & CMU) | Unified environment for evaluating agents on desktop OS tasks | [arXiv:2404.07972](https://arxiv.org/abs/2404.07972) | | **2025** | *Agent2Agent (A2A) Protocol Specification* | Google Cloud & DeepMind Working Group | Standardized inter-agent negotiation and capability discovery | [Google Tech Report](https://developers.google.com) | | **2026** | *OWASP Top 10 for Agentic Applications Standard (ASI Top 10)* | OWASP Foundation Working Group | Industry-wide risk taxonomy and mitigation framework | [OWASP Standard](https://genai.owasp.org/) | [Back to Top](#nexum--the-agentic-universe) --- ## 12. Multi-Language Agent Framework Matrix *Quick-reference ecosystem index categorized by programming language.* | Language | Primary Frameworks & SDKs | Key Features & Focus | | :--- | :--- | :--- | | **Python** | LangGraph, CrewAI, PydanticAI V2, Swarms, AutoGen v0.4, Smolagents, Camel-AI | Stateful graphs, role-based swarms, type validation, micro-VMs | | **TypeScript / Node** | Mastra, LangChain.js, AutoKitteh, Flowise, Midscene.js | Full-stack JS/TS native, React integration, workflow canvases | | **C# / .NET** | Microsoft Agent Framework (MAF 1.0), Semantic Kernel | Enterprise Azure/Office 365, declarative YAML, high throughput | | **Go** | Google ADK 2.0 (Go), LangChainGo, E2B Go Client | High concurrency, microservices orchestration, low memory | | **Rust** | Rig, AutoGPT Rust Core, Llama.cpp Rust Bindings | Maximum performance, zero-cost abstractions, edge agent servers | | **Java / Kotlin** | LangChain4j, Semantic Kernel Java, Google ADK Java | Enterprise JVM backends, Spring AI integration | [Back to Top](#nexum--the-agentic-universe) --- ## 13. The NEXUM Standard ### *How we curate the world's best list* To be included in NEXUM, a project must undergo the **Lead Designer's Vetting Process**: ```text ┌─────────────────────────┐ ┌─────────────────────────┐ │ 1. Autonomy Tier │ ──► │ 2. 2026 Resilience │ │ Multi-step reasoning │ │ MCP / A2A / AP2 / x402 │ └─────────────────────────┘ └─────────────────────────┘ │ │ ▼ ▼ ┌─────────────────────────┐ ┌─────────────────────────┐ │ 3. Community Velocity │ ──► │ 4. Operational Safety │ │ Active code & tests │ │ OWASP ASI Top 10 Ready │ └─────────────────────────┘ └─────────────────────────┘ ``` 1. **Autonomy Tier**: Must demonstrate multi-step reasoning, goal decomposition, and task execution without human micro-management. 2. **2026 Resilience**: Must support modern protocols (MCP, A2A, ACP/AP2/x402, AG-UI, or VLA). 3. **Community Velocity**: Active development with high-fidelity documentation, reliable test suites, and proven production usage. 4. **Operational Safety**: Built-in guardrails, telemetry hooks, or compatibility with the OWASP ASI Top 10 and other Guardian systems. --- ## 14. Join the Vanguard We welcome contributions from the open-source community to keep NEXUM the premier reference for autonomous systems. ### 14.1. Contributing Workflow 1. **Fork** the repository: [github.com/HA2345567/awesome-autonomus-ai-agents](https://github.com/HA2345567/awesome-autonomus-ai-agents) 2. **Add Your Tool** in alphabetical or relevance order within the matching architectural layer: ```markdown - **[Tool Name](https://link.to.tool)** `License / Badge` `Tag` — Authoritative description highlighting 2026 context and standout capabilities. ``` 3. **Validate**: Ensure links are live, formatting matches the surrounding section, and the tool satisfies the [NEXUM Standard](#13-the-nexum-standard). 4. **Open a Pull Request**: Submit your PR with a clear summary of the addition. ### 14.2. Contributors <div align="center"> <a href="https://github.com/HA2345567/awesome-autonomus-ai-agents/graphs/contributors"> <img src="https://contrib.rocks/image?repo=HA2345567/awesome-autonomus-ai-agents" alt="Contributors" /> </a> *Thank you to all the brilliant engineers, researchers, and creators shaping the agentic frontier.* </div> ### 14.3. Star History <div align="center"> [![Star History Chart](https://api.star-history.com/svg?repos=HA2345567/awesome-autonomus-ai-agents&type=Date)](https://star-history.com/#HA2345567/awesome-autonomus-ai-agents&Date) </div> --- <div align="center"> ### Support the Autonomous Revolution If you find NEXUM valuable, please consider **starring the repository** and sharing it with your network! [![Star on GitHub](https://img.shields.io/badge/Star%20Repo-★%20NEXUM-0969da.svg?style=flat-square&logo=github)](https://github.com/HA2345567/awesome-autonomus-ai-agents) <br/> *Curated with precision by **NEXUM** • Continuously Updated for 2026 & Beyond* [Back to Top](#nexum--the-agentic-universe) </div>

awesome-autonomus-ai-agents

内容

连接信息

猜你喜欢

awesome-autonomus-ai-agents

微信扫一扫分享

需要认证

内容

连接信息

猜你喜欢