Supplemental material for Private Agent Systems by Daniele Verì: a manual for designing private, observable, and governed AI agent systems.
Book • Companion repositories • Supplemental chapters • Contents • Repository structure • Errata • Source updates • Contributing • License
Amazon: US | UK | CA | AU | DE | FR | ES | IT | NL | SE | PL | MX | JP | IN
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.
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
Building a Governed LLM Wiki as a Second Brain 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/ 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 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 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 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.
This repository is the stable place for material that may change after publication:
templates/: fill-in versions of the reusable artifacts from Appendix B (route cards, eval cases, data flow, policy gates, and more)chapters/: supplemental chapters published after the first editionevidence/: sanitized public evidence tables referenced by supplemental chaptersstarter-kit/: runnable second-brain vault skeleton for the governed LLM Wiki chaptersource-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.
.
├── README.md This file.
├── LICENSE Licensing terms (CC0 for templates/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
│ ├── governing-context-budgets.md
│ └── how-to-trust-a-liar.md
├── evidence/
│ └── token-saving-public-results.md
├── starter-kit/ Portable Claude Code and Codex second-brain skeleton.
├── tests/ Lightweight checks for public evidence artifacts.
├── tools/ Utility scripts for reproducible evidence generation.
└── 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 templates + source register.
├── 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 for what each template is and which chapter it comes from.
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.
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.
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.mdwith 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.
The fill-in templates in templates/, the portable files in
starter-kit/, and
source-register.md are dedicated to the public domain
under CC0 1.0 (see 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, and the explanatory text in these README files, is the author's reserved work, provided for reading and reference only.