Governor’s Log
What we are learning about AI agents, their architectures, and the systems required to understand and govern them.
Sixteen Approvals for a Copy
How a coding agent and an AI architect turned a reversible copy into a gated migration, and why a human approval every four minutes is not oversight.
A Valve Is Not a Read
Integrating Sentience Governor into Antibody, a self-healing building of 14 agents and a coordinator on Flower, and what physical actions exposed.
The Price of Remembering
A hackathon build on Claude, Pydantic AI and Sentience Governor that kept an agent’s mission through context compaction, and measured the token price of remembering.
Parseable Is Not Valid
Why a Sentience Governor profile that parsed was not valid, what fail-open must preserve, how a fingerprint becomes identity, and what a tests-first corpus proved.
The Vocabulary of Agent Governance Is Still Being Written
Why Sentience published a 50-entry glossary of AI agent and runtime governance terms, how each definition was checked, and where the definitions stop.
What Does an Agent Action Actually Mean, and Who Governs It?
Sentience Governor 0.3.2: per-agent policy resolution that stays stable for a session, and semantic classification of Claude Code shell actions by their effects.
Sentience Governor for Pydantic AI
pydantic-ai-governor 0.1.0 is live. Attach one capability and a Pydantic AI run carries a record of what the agent declared and what it dispatched.
What's Your Claude Up To?
Sentience Governor v0.3.1.1 adds an evidence path to its retrospective review of Claude Code history, and what four months of our own transcripts showed.
Nothing Obviously Broke
Claude Code restarts produced valid but empty Sentience Agent Execution Records, exposing how correct evidence can still mislead a runtime governance system.
The Fix That Grew
A stale Claude Code hook left Sentience Governor without execution records for five weeks. Why the first repair design solved the wrong problem.
The Policy Was Right. The Run Was Wrong.
A LangGraph false-positive fix shows why a governance control can have the right rule and still be attached to the wrong execution.