Start with authority
A coding agent needs to know which repository it can change, which tools it can invoke, and which actions require a human decision. A prompt can describe these boundaries; the execution layer must enforce them.
Make a run inspectable
Record the input, tool call, result, file changes, and a reason for stopping. Treat a model response as a proposal, not proof that a task succeeded. An independent build or test provides stronger evidence.
A useful run envelope includes:
interface RunBudget {
maxToolCalls: number;
maxDurationMs: number;
maxCostUsd: number;
}The next experiment
Start with one bounded repository task. Compare the proposed plan with the resulting diff and test output. Track interventions alongside completion rate.
This connects to replayable failure scenarios and the AI Development Engine.
Open question
Which decisions deserve a human checkpoint, and which can be safely validated by a deterministic rule?