Per-action approval became per-run, and Agile already had the words for it

I went back and looked at how often my approval of an agent’s recommendation actually changed the recommendation.

Almost never. The overwhelming majority of the per-item approvals I was issuing did exactly one thing: ratify what the agent already proposed. I was a rubber stamp with a pulse, and the stamping was the slowest part of the whole system.

i.The trap in human-in-the-loop autonomy

That is the trap in human-in-the-loop autonomy, and it is easy to walk into from the right instinct. You want a human on the irreversible acts, so you put a gate on the agent’s output. Then you notice the agent is good, so the gate becomes a formality. But a formality you cannot remove is not a safety control. It is latency wearing a safety control’s clothes.

The fix was to stop gating actions and start gating runs.

ii.Gate the run, not the action

Here is the shape. The human approves a plan up front: what the agent is about to do, the constraints it must hold, what done looks like. Then the agent runs the entire unit of work alone. Every reversible step happens ungated. Every irreversible step (publish, send, spend, delete, or locking something into the source of truth) does not fire. It gets staged as a draft or a held flip. When the run is done, the human reviews the whole batch at once and approves or redlines the lot.

Two gates per run, not one per action. The human touches the work twice: before, to point it, and after, to check it. In between, the agent runs the cleared ground.

This only works because of one rule that never bends: the agent never approves its own irreversible work and then ships it. Reversible work is free because git is the floor under it; anything you can undo with a revert does not need a human standing over it. The gate is reserved for the acts you cannot take back.

iii.I had reinvented a sprint

Now the part that surprised me. When I wrote this model down and looked at it, I realized it was not new. I had reinvented a sprint.

The plan I approve before the run is sprint planning. The batch review after is the sprint review. The handoff I write so the next session resumes cleanly is the retro. The decision log is a stack of architecture decision records. The backlog of work the agent drains is the product backlog. The first sprint, the one I am running right now, is the one where you clear the backlog.

I did not set out to do Scrum. The constraints of trustworthy autonomy pushed me into the same shape that constraints of trustworthy teams pushed an industry into decades ago. That convergence is the signal that the shape is load-bearing, not fashionable.

iv.Keep the spine, leave the ceremony

To be clear about what I did not adopt: no agent standups, no story points, no fixed-length sprints. That is the ceremony, and the ceremony is the part of Agile that calcified into theater. I took the spine, which is short feedback loops with the human at planning and review, and left the ritual.

The portable principle: once an agent can run a unit of work without you, move your gate from the action to the run. Approve the plan, let it work, review the batch. Keep your hand on exactly one thing, the irreversible act, and let everything reversible run free. Conflate the two and you turn yourself back into a clerk on work you already trust.

Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me. Model provenance: Claude Code (Claude Opus and Sonnet)