AI Agent Architecture

The platform is what every agent shares and none of them owns.

Atlas, Meridian Supply Co.'s customer-operations agent, was certified by Book 6. Then Sales asked for a second agent, and the estimate was to copy six thousand lines. This book extracts what every agent needs and none should own into an enterprise platform — and registers the second agent in thirty-six lines.

Every chapter is a lab on the companion repository — one dependency, fully offline with a scripted mock, every listing printed from a verified line range, every command paired with its expected output, and the book's own modules and demos printed in full in Appendix E. Five moves per chapter: Run the demo, Read the listing, Break it with the chapter's planted failure, Fix it with the design move, Prove it with a test that turns green. You will:

• Extract the model, memory, tools, data, core, and eval gate into services, inject them into an unchanged Atlas, and measure the second agent's onboarding cost
• Name six layers, write the one import rule, and turn it into a lint that fails the build — with an exception table printed beside the violations
• Make departments tenants with their own stores and budgets, and test the cross-tenant read that is denied and the neighbour that never notices
• Build an identity fabric where only a person can start a chain, delegation only narrows, and every gateway call records who authorised it
• Serve capacity by share, shed by priority, and turn the meter into a bill with the fixed cost on its own line
• Move versions through dev, staging, and prod by gates, and roll back a hotfix that passed every suite when the canary reads 53 percent
• Tag every knowledge field with a class, mask personal data before the brief and again at the model's door, and count what the door catches
• Wrap a 2011 returns system in a facade, shift traffic by parity, and find the rule nobody wrote down on the thirtieth day
• Guard the architecture with nine fitness functions, keep the list of what they cannot measure, and plan the next generation as decision records
• Generate the reference architecture and a three-year roadmap from the code and the records, so that neither can drift

Every chapter carries a research lineage (platform engineering and Team Topologies, Dijkstra's layers and Clean Architecture, SaaS tenancy, OAuth and zero trust, fair queueing and FinOps, continuous delivery and canaries, data lineage and the GDPR, the strangler fig and the anti-corruption layer, evolutionary architecture, living documentation), a five-item failure catalog, an applied deep-dive, and exercises in three tiers. Ten figures, 351 tests, sixty-two decision records, one running system that became a tenant of the platform built under it.

Who it's for: platform and enterprise architects, CTOs, and the engineers who will run the second, third, and tenth agent on shared infrastructure. Assumes Book 1 and a skim of Books 2–6; basic Python; no framework, no GPU.

The AI Agent Architecture Series is the architect's track: one architectural layer per book, on one running system the reader refactors and grows by hand. This is Book 7, The Enterprise Platform — the last.