Most agent supervision is a gate at the end of a road. The agent works alone for an hour, presents a finished artifact, and a reviewer human or machine accepts or rejects it. This is terminal review, and it has a structural flaw: by the time the gate sees anything, every mistake has already been committed. A wrong assumption in minute two has spent fifty-eight minutes compounding. Rejection at the gate doesn't undo the work; it discards it. The agent starts over, often with no better information than it had the first time.
Human teams don't work this way, and it's worth asking why. A senior engineer pairing with a junior one doesn't sit silent until the pull request lands. She glances over occasionally. She says "check the timezone on that" in passing. Most of her interventions are small, early, and forgettable which is exactly why they work. The expensive review at the end exists too, but it catches little, because the cheap corrections along the way caught nearly everything.
A verifier inside the loop
This early design placed a verifier inside the agent's working loop, watching work in progress rather than waiting for the final transcript. When it saw drift beginning—an assumption that contradicted the task, a skipped check, a repeated dead end—it could correct the work in-channel. We called the pattern the silent voice: unobtrusive to the user, available to the working agent.
The design bet is that the cost of an intervention should match the size of the error. Most errors in long agent runs are small when they're born. A steering nudge at that moment costs almost nothing. The same error at terminal review costs the whole run.
Hard stops still exist
Steering does not replace stopping; it makes stopping meaningful. In this design, hard stops were reserved for bright lines: movement of money outside an approved leash, an irreversible external effect, or a stated user boundary. The broader lesson survives even as the implementation changes: frequent low-value stops teach people to route around a safeguard.
The second job: adjudication
The prototype also used a terminal evidence gate. It asked what in the record proved completion rather than whether the result merely looked finished. That specific gate has since been retired; the durable lesson is to keep claims attached to artifacts, tool results, and checks.
This is an archived design account, not a description of a separate agent you can select today. Current Neo uses integrated evidence-sensitive reasoning, while product completion still depends on the evidence appropriate to the job.
What changed when we retired gates-only supervision
We compared gates-only supervision with this steering pattern qualitatively in internal work. We are not presenting that experience as a benchmark. It informed the direction toward catching drift while correction is still cheap and keeping hard boundaries structural.
The lesson we keep relearning: supervision is not a checkpoint, it's a presence. Gates tell an agent when it has failed. A voice in the loop keeps it from failing. You need both but if you only have gates, you're grading autopsies.