Vercel's agent stopped at the line and asked a human to cross it

On October 23, Vercel shipped a production agent that stops at its permission boundary and asks. The grant stays visible. The human decides whether to extend it. That behavior looks like a constraint on the agent, but the design consequence falls on the reviewer: they now hold a request with a clear object and a clear cost, instead of a fait accompli they can only ratify. That shift in what the human holds changes what the approval means.

The same day, DoorDash's reviewers approved or rejected 35,000 code changes for $4.79 each, working from summaries the agents wrote. The reviewer never inspected the code. The summary was the interface, and the accuracy of that summary was the design problem no one had named yet. Atlas, a tool for auditing long agent runs, showed the same gap from the other side: the agent ran, the human arrived after, and the review screen had to carry the full weight of a session the human did not witness. Vercel's stop-and-ask behavior threads these two problems together. The agent produces a legible request rather than a completed act, and the human reads a summary of what is being asked rather than what was done.

The arXiv paper on agent groups adds a harder version of the same problem. High confidence from a group of agents can mean they all copied the first one. The review surface showed agreement; it did not show where agreement came from. A reviewer reading that surface had no way to distinguish genuine convergence from cascading repetition. The summary concealed the structure of the evidence, and the approval concealed its own fragility.

The RCT on AI-assisted writing found that a reflection pause before generation lifted output quality. The scaffold changed what the writer knew before the AI started, and that prior position made reviewing the output possible. The same logic applies here. ABRA binds each agent-generated panel to the agent that built it, so the human can see provenance before touching the surface. The HCI workspace research found that the workspace holds the knowledge, and a team that designs the agent before the workspace leaves the human approving without enough to judge. Each of these findings points at the same moment: the one before the human decides, and how much the design has placed in view by then.

The practical instruction is narrow. Build the request surface before you build the capability. For each action your agent can take, design the object the human reads when the agent stops and asks: name the action, name the scope, name what the agent cannot see from where it stands. Then design the review surface for actions the agent completed without stopping: show the provenance of each claim, not the confidence score that summarizes it. Vercel's stop behavior is the easy half. The summary the reviewer reads is the half most product teams have not built yet.