Agentic Loops

Drive a concrete task to completion through short, auditable cycles. Keep every loop human-started, bounded, reversible where possible, and subject to ordinary approval rules.

Execute the loop

1. Record the goal, observable acceptance criteria, constraints, current state, and budgets for iterations, elapsed time, cost, and tool calls.
2. Inspect the smallest evidence surface that can reveal the next high-value action.
3. Choose one hypothesis or action. Predict the signal that would count as progress before acting.
4. Act within scope. Keep the change small enough to attribute the observed result.
5. Verify against the real acceptance criteria, not a proxy. Record action, observation, verification, artifacts, spend, and whether measurable progress occurred.
6. Adapt the next action from the new evidence. Do not repeat an unchanged failed action.
7. Stop immediately when every acceptance criterion passes, an approval boundary is reached, the user cancels, the iteration/time/tool/cost budget is reached, or two consecutive iterations show no progress without a new hypothesis. A zero cost cap permits zero-spend work but is reached by any positive recorded spend.

Maintain state