notes · working out loud

Notes on agents that act, and how to trust them.

Shorter than a paper, faster than review. The thinking behind the research, audit trails, multi-agent collusion, and catching agents that quietly misbehave, written while it's still forming.

14 Jul 20265 mindraft
Your agent's logs are not evidence
An agent that writes its own audit trail can rewrite it. That is the whole problem.
audittamper-evidencesasanaRead ↗
9 Jul 20266 mindraft
Agents told not to collude will collude anyway
A 'no side-channels' policy doesn't stop coordination. Sometimes it just moves it.
multi-agentcollusionfacadeRead ↗

Your agent's logs are not evidence

Agents told not to collude will collude anyway