Private-Agent-Systems

# Private Agent Systems: resources Supplemental material for *Private Agent Systems* by Daniele Verì: a manual for designing private, observable, and governed AI agent systems. [Book](#book) • [Companion repositories](#companion-repositories) • [Supplemental chapters](#supplemental-chapters) • [Contents](#contents) • [Repository structure](#repository-structure) • [Errata](#errata) • [Source updates](#source-updates) • [Contributing](#contributing) • [License](#license) --- ## Book <p align="center"> <a href="https://www.amazon.com/dp/B0H4RQMJG3"> <img src="https://m.media-amazon.com/images/I/416Yvhd0zVL.jpg" alt="Private Agent Systems cover" width="320"> </a> </p> <p align="center"> <strong>Amazon:</strong> <a href="https://www.amazon.com/dp/B0H4RQMJG3">US</a> | <a href="https://www.amazon.co.uk/dp/B0H4RQMJG3">UK</a> | <a href="https://www.amazon.ca/dp/B0H4RQMJG3">CA</a> | <a href="https://www.amazon.com.au/dp/B0H4RQMJG3">AU</a> | <a href="https://www.amazon.de/dp/B0H4RQMJG3">DE</a> | <a href="https://www.amazon.fr/dp/B0H4RQMJG3">FR</a> | <a href="https://www.amazon.es/dp/B0H4RQMJG3">ES</a> | <a href="https://www.amazon.it/dp/B0H4RK4HY6">IT</a> | <a href="https://www.amazon.nl/dp/B0H4RQMJG3">NL</a> | <a href="https://www.amazon.se/dp/B0H4RQMJG3">SE</a> | <a href="https://www.amazon.pl/dp/B0H4RQMJG3">PL</a> | <a href="https://www.amazon.com.mx/dp/B0H4RQMJG3">MX</a> | <a href="https://www.amazon.co.jp/dp/B0H4RQMJG3">JP</a> | <a href="https://www.amazon.in/dp/B0H4RQMJG3">IN</a> </p> **Private Agent Systems** *Architecture, privacy, observability, and governance for local, hybrid, and cloud AI agents* Daniele Verì The book treats agents as production software systems, not prompt experiments. It focuses on the harness around the model: identity, data classification, retrieval, model routing, tool permissions, memory, observability, evaluation, privacy controls, and operational ownership, and on turning privacy, consent, security, tool governance, and rollout into runtime decisions that can be inspected and reviewed. Two first-party case studies, **Amber** (governable Hybrid GraphRAG) and **Relic** (runtime governance, provenance, consent, and memory control), ground the architecture in concrete systems. - **English edition (canonical):** *Private Agent Systems*: first edition - **Italian edition:** *Sistemi agentici privati* *Daniele Verì builds private, observable AI systems, where cognitive science meets engineering. He is the author of Amber (governable Hybrid GraphRAG) and Relic (runtime governance for agent memory and consent): the two case studies in this book.* --- ## Companion repositories The two first-party case studies from the book are open source: - **Amber** (governable Hybrid GraphRAG): https://github.com/yuzushi-dev/Amber - **Relic** (runtime governance, provenance, consent, memory control): https://github.com/yuzushi-dev/Relic --- ## Supplemental chapters **[Building a Governed LLM Wiki as a Second Brain](chapters/building-a-governed-llm-wiki.md)** is a practical guide to maintaining a local, source-linked knowledge base with Claude Code and Codex. It covers the knowledge lifecycle, operational memory, filesystem MCP wiring, provenance, corrections, concurrent writes, security, and end-to-end acceptance tests. The accompanying [`starter-kit/`](starter-kit/) directory is a portable second-brain starter with matching `CLAUDE.md`, `AGENTS.md`, `00-system/`, source folders, wiki folders, and client configuration files. **[Governing Context Budgets in Private Agent Systems](chapters/governing-context-budgets.md)** shows how to treat context budget as a runtime governance surface. It covers evidence classes for token-saving claims, dated local audit results, tool selection by purpose, trace requirements, approval-policy implications, and the limits of prompt-token estimates. **[Agentic Loops in Private Agent Systems](chapters/agentic-loops.md)** shows how to design recurring agent work as an operating contract. It covers loop admission tests, route cards, threat models, maker/checker separation, context budgets, traces, human gates, downgrade paths, and staged adoption. **[How to Trust a Liar: Verifying Model Claims in Private Agent Systems](chapters/how-to-trust-a-liar.md)** shows how to build a runtime that treats every model claim as unverified until checked. It covers confabulation versus lying, maker/checker separation, repetition metrics, a source register that keeps verified, failed, and not-checked outcomes distinct, leakage denylists, and the human content-quality gate. **[From Loop Engineering to Graph Engineering: Agent Topology as a Review Surface](chapters/from-loop-engineering-to-graph-engineering.md)** treats agent topology as a set of contracts between bounded loops rather than a diagram. It covers the admission test for graph review, the three surfaces (knowledge relationships, execution transitions, decision lineage), declared checks with evidence and owners, consumed-result binding as the test for a real edge, failure and privacy review, and a worked example from Amber's graph write path. The accompanying packet is machine-checkable: [`templates/topology-audit-schema.json`](templates/topology-audit-schema.json) declares the registers, evidence entries, and checks; [`tools/topology_audit_validate.py`](tools/topology_audit_validate.py) runs the structural rules; and [`evidence/`](evidence/) holds one passing packet plus negative packets for missing fields, invalid status values, duplicate identifiers, dangling references, route gaps, and unmapped requirements. --- ## Contents This repository is the stable place for material that may change after publication: - [`templates/`](templates/): fill-in versions of the reusable artifacts from Appendix B (route cards, eval cases, data flow, policy gates, and more) - [`chapters/`](chapters/): supplemental chapters published after the first edition - [`evidence/`](evidence/): sanitized public evidence tables referenced by supplemental chapters - [`starter-kit/`](starter-kit/): runnable second-brain vault skeleton for the governed LLM Wiki chapter - [`source-register.md`](source-register.md): the dated source register from Appendix B, kept current here as sources change - Companion links for Amber and Relic (above) - Errata and clarifications (added as they are reported) The print book, ebook, and Kindle edition may refer readers here when a resource is easier to maintain outside the manuscript. --- ## Repository structure ``` . ├── README.md This file. ├── LICENSE Licensing terms (CC0 for code, templates, evidence, register; book text reserved). ├── CITATION.cff How to cite the book. ├── source-register.md Dated source register (Appendix B), updated as sources change. ├── assets/ │ └── cover.jpg Full-size book cover. ├── chapters/ │ ├── agentic-loops.md │ ├── building-a-governed-llm-wiki.md │ ├── from-loop-engineering-to-graph-engineering.md │ ├── governing-context-budgets.md │ └── how-to-trust-a-liar.md ├── evidence/ JSON packets are CC0; the results write-up is reserved. │ ├── token-saving-public-results.md │ ├── topology-audit-positive.json │ └── topology-audit-negative-*.json Packets that must fail validation. ├── starter-kit/ Portable Claude Code and Codex second-brain skeleton. ├── tests/ Lightweight checks for public evidence artifacts (CC0). ├── tools/ Utility scripts for reproducible evidence generation (CC0). └── templates/ Fill-in operational artifacts from Appendix B (CC0). ├── README.md Index of all templates, mapped to their book chapters. ├── LICENSE CC0 1.0 dedication for the code, templates, evidence, and register. ├── topology-audit-schema.json ├── route-card.yaml ├── data-class-table.yaml ├── data-flow.yaml ├── provider-review-sheet.yaml ├── source-revalidation-row.yaml ├── policy-gate-matrix.yaml ├── agent-suitability-record.yaml ├── open-source-harness-review-card.yaml ├── local-inference-readiness-card.yaml ├── model-adaptation-decision-record.yaml ├── context-budget-worksheet.yaml ├── eval-packet.yaml ├── eval-stack-card.yaml ├── agent-eval-case.yaml ├── rag-eval-case.yaml ├── prompt-regression-case.yaml ├── trace-linked-evaluation-record.yaml └── incident-report.yaml ``` See [`templates/README.md`](templates/README.md) for what each template is and which chapter it comes from. --- ## Errata If you find a technical error, outdated reference, broken link, or unclear passage, open an issue in this repository. Include the edition, format, chapter, section, and enough context to reproduce the problem. For privacy, security, legal, or procurement decisions, treat the book as technical education. Re-check current vendor documentation, laws, contracts, and organizational policy before using any pattern in production. --- ## Source updates The book relies on time-sensitive material: model cards, provider documentation, standards, security guidance, open-source repositories, and product terms. When a source changes after publication, record the updated link, date, and impact on the relevant claim in [`source-register.md`](source-register.md). --- ## Contributing Contributions are limited to keeping the supplemental material correct and current: - **Errata**: open an issue (see above). Pull requests that fix a typo, broken link, or outdated reference in this repository are welcome. - **Source updates**: if a cited source moved or changed, propose an update to [`source-register.md`](source-register.md) with the new link and the date you checked it. - **Templates**: the templates are intentionally minimal patterns. Suggestions that keep them general and chapter-faithful are welcome; project-specific variants belong in your own repository. Please do not open pull requests that reproduce or redistribute the book text. --- ## License The fill-in templates in [`templates/`](templates/), the portable files in [`starter-kit/`](starter-kit/), the utility scripts in [`tools/`](tools/), the checks in [`tests/`](tests/), the machine-readable artifacts in [`evidence/`](evidence/), and [`source-register.md`](source-register.md) are dedicated to the public domain under **CC0 1.0** (see [`templates/LICENSE`](templates/LICENSE)). Copy, modify, and use them freely, including commercially, with no attribution required. Everything else, including the prose of the book *Private Agent Systems*, the supplemental chapter prose, the cover image, the written results commentary in [`evidence/token-saving-public-results.md`](evidence/token-saving-public-results.md), and the explanatory text in these README files, is the author's reserved work, provided for reading and reference only.