Whitepaper

Governed agents: a runtime for policy-bounded, auditable AI agents

How Camaan's agent runtime puts every tool call through one policy-checked path and signs each outcome into a tamper-evident ledger. The paper gives the design, the invariants and where each is enforced, and what the authors do and do not claim.

Every tool call, whatever its origin, passes one ordered path of checks before it runs.
Approvals are bound to the exact call, and the person who raises a change cannot approve it.
Each decision is recorded in a signed, hash-chained ledger; signing a record costs about 19 microseconds.
Credentials stay in an encrypted vault and never enter the model's context; interrupted turns resume without repeating finished calls.