The Keyboard Belongs To The Human At Risk

Autonomy Ends Where Uncertainty Becomes Expensive

A coding agent earns the keyboard by staying inside a narrow band of risk, and it keeps that keyboard only while the band holds. The band is defined by three qualities that matter more than raw capability: the task is specified tightly enough to audit, the change is cheap to reverse, and the agent can explain its own edits without guessing. When any one of those qualities slips, the cost of a wrong move starts to exceed the cost of a human interruption.

The moment of handback is not a failure of the agent. It is a design decision made in advance, before the pressure of a running test suite or a looming release makes the choice feel like surrender. Teams that write the boundary down treat the handback as a normal part of the workflow. Teams that leave it implicit end up debating the boundary in the worst possible moment, with half-written changes on the table and no clean record of intent.

Three questions sit at the center of that boundary. Is the correct outcome described precisely enough that two engineers would agree on it? Can the agent point to the evidence that its change does the right thing? Is the blast radius small enough that an incorrect edit costs an hour, not a sprint? A confident answer to all three keeps the agent typing. A hesitation on any one is the signal.

Side-By-Side Criteria Separate Drafting From Deciding

Read the two columns together and the split becomes mechanical. On one side sit situations where the agent acts as a careful drafter inside a known frame. On the other sit situations where it is asked to decide something the frame does not cover.

• Keep the keyboard when the change is local, the acceptance criteria are written down, and existing tests describe the behavior being preserved.
• Keep the keyboard when the agent can name the reason behind each edit and the diff can be reviewed in a single sitting.
• Keep the keyboard when the environment is disposable, such as a scratch branch or a sandboxed service that no customer touches.
• Keep the keyboard when repeated runs converge on the same result and variation stays inside an agreed tolerance.

• Hand it back when the request contains a trade-off, such as speed against safety or storage against simplicity, and no preference has been recorded.
• Hand it back when two reasonable engineers could read the same requirement and design different systems.
• Hand it back when a change touches credentials, money, personal data, deletion, or an external contract that customers depend on.
• Hand it back when the agent cannot produce evidence for its claim, or when gathering that evidence would take longer than a human decision.

Notice that none of these criteria concern model quality. A stronger model widens the first column without emptying the second one. The questions about stakes, ambiguity, and reversibility survive every generation of model, and they remain the right ones to ask before the first edit is made.

Escalation Triggers Show Up Before Tests Fail

The strongest signals arrive early, usually in the shape of conversation rather than error output. The agent starts asking clarifying questions it did not ask before. It proposes three architectures in one reply instead of committing to one. It hedges its language, marks its own plan as a guess, or repeats a requirement back with a qualifier attached. These are not stylistic quirks; they are the model reporting that the instructions no longer pin down the answer.

Operational signals matter just as much. Retry counts rise, the same edit is applied and reverted twice, a test fails for a reason the agent admits it does not understand, or a dependency changes underneath the plan. Any one of those is a reason to pause, and a cluster of them is a reason to stop. The temptation is to treat a red test as a puzzle to be solved at all costs. That temptation is exactly where agents quietly break working code to make a number go green.

Set the threshold before the session starts. A practical rule is that an agent may spend a bounded amount of effort exploring, and then must report its findings and its proposed next step rather than continue. Bounded effort keeps the cost of a wrong path predictable, and it turns the handback into a scheduled event instead of an emergency.

Handoffs Need Evidence, Not Apologies Or Excuses

A good handback is a small technical document. It states the goal in one sentence, lists the changes already made, names the tests that passed and the ones that did not run, and describes the decision the human is being asked to make. The agent should offer a recommendation, because the person receiving the handoff is usually capable of reading one and correcting it faster than inventing one from scratch.

An apology is not evidence. Neither is a long narrative about effort spent. The reviewer needs the shape of the change, the confidence behind each part, and the exact point where confidence ran out. When that packet is complete, the human can resume in minutes. When it is missing, the human has to reconstruct the session, and the value the agent created evaporates into archaeology.

Tooling choices shape this packet more than people expect. A quality-first relay such as BestRelay forwards each request to the official model named in it, with no substitution, so behavior in one session matches behavior in the next. Teams running Claude Code, Codex, Gemini CLI, or any client built on the OpenAI SDK keep that routing discipline, and pay-as-you-go billing follows the token counts the provider reports, with cache reads and cache writes priced separately. Stable routing makes regression reports trustworthy too, since a change in behavior can be attributed to the code rather than to an undocumented model swap. Model names such as Claude Sonnet 5.5 stay exactly as specified in the request.

Good Handoffs Keep The Human Loop Short

The purpose of drawing the line is speed, not caution for its own sake. An agent that keeps typing through uncertainty is slower in aggregate than one that stops early, because the human eventually has to untangle decisions that were never theirs to make. A crisp handback preserves momentum: the human decides one thing, the agent resumes with a narrower brief, and the work continues with the risky part behind it.

Teams that practice this well treat the boundary as a shared artifact. They keep a short list of domains where the agent acts alone, a short list where it acts only with approval, and a short list where it advises and nothing more. They revisit the lists after every incident, because incidents are the only honest test of whether the boundary was drawn in the right place.

There is also a human courtesy built into the practice. A person interrupted with a clear question and a clean summary feels respected, while a person who discovers a silent rewrite of a critical path at midnight does not. Members who invite colleagues earn a referral commission when those colleagues top up. Create a BestRelay account and run the next session against a stable endpoint.