Operating AI in production.
Evergreen chapters for people putting AI into production — why governance matters, how evidence differs from logs, how to think about agent identity, and why production AI needs a different operational model than the software before it.
AI agents are becoming production software. Production software needs governance.
Start here
The core ideasReplay
AI systems rarely fail the same way twice. Reconstructing what happened is often harder than fixing it.
Read the chapterEvidence
The day someone stops trusting your word, logs become surprisingly useless.
Read the chapterReceipts
Every meaningful action deserves a small, signed record. Most systems still leave prose.
Read the chapterAuthorities
On whose authority did the agent act? "The model" is not a defensible answer.
Read the chapterGovern
Rules an AI agent will actually follow, not policies buried in a wiki.
Read the chapterRun
A governed action needs a point, before execution, where policy is evaluated and the outcome recorded.
Read the chapterProve
Anyone should be able to verify a past decision without your help.
Read the chapterAlso in the guide
Implementation ideas- Policies
How rules compile into portable, versioned bundles.
- Identity
Agents aren't users and aren't services — how to think about non-human callers.
- Delegation
Scoped, time-boxed authority passed from one party to another.
- Verification
How a stranger checks a receipt bundle without access to your system.
- Kill switch
Stopping future governed consequences in a selected class without stopping everything.
- Local-first
Why keeping credentials and history on-device changes the trust model.
- Governance vs security
Security stops attackers. Governance decides what your own agents may do.
- Glossary
The consequence architecture: Consequence, Decision, Authority, Attempt, Effect, and Evidence.
Looking for the how, not the why? Go to Docs.