Autonomous livecoding agents must apply stricter cadence discipline than copilot mode, defaulting one tier smaller

In copilot mode the human’s editor is the source of truth and the agent proposes one patch-tier diff at a time, waiting for the human to save. In autonomous mode the agent saves directly, so cadence discipline is stricter: bias one tier smaller than in copilot (micro/patch), space edits by cycle counts, and batch nothing — land one concept-id, wait a phrase, then reassess. An autonomous agent editing every cycle is the ‘runaway’ failure the livecoding community fears most. Note: the agent currently cannot perceive results (no L3 perception bridge) so reassessment is based on cycle count, not listening.

Examples

Copilot proposes a patch; human decides when to save. Autonomous agent makes a micro/param change, waits 4 cycles, then makes the next change.

Assessment

Explain why autonomous livecoding agents must apply stricter cadence discipline than copilot mode, and state what ‘reassess after a phrase’ means given the current perception constraint.