Regimes — auditable improvement loops via event-sourced agents
Most agent improvement loops live outside the agent—failures stay unlogged, replays are impossible, decisions hide in side databases. Regimes embeds the loop into the agent's runtime itself.
• Agent state is a deterministic projection of an append-only event log
• Every failure, patch, gate decision, and promotion is recorded as an event
• Runs replay exactly from their log; candidate patches scope to typed pipeline seams
• Improvement becomes auditable, first-class workflow, not external scaffolding
Demonstrated on LongMemEval with ActiveGraph runtime.