EAIR found AI teams overstep even when each member stays in bounds

On October 14, an EAIR paper described a pattern that will be familiar to anyone who has watched a project budget slip: each agent on the team operated inside its assigned scope, and the team crossed the line anyway. Budget-scoping and reach-preview are both designed around a single agent and a single human reviewing that agent's share. The group's running total, the number that actually determines whether the work stays within bounds, exists nowhere a reviewer can see it.

Three other papers published the same day show the same shape from different angles. The arXiv study of failing AI teammates found that the human handed the work off and received no signal when one member of the group went wrong. The work looked like progress. The paper on planted ads in search results found that the human approved the task at the start and reviewed the result at the end, and the middle, what the agent read, what it was told inside the pages it visited, stayed invisible. The expertise study found that experts caught errors because they reviewed behavior differently, not because the product gave them better information.

The common object in all four cases is the handoff screen. That screen is where the reviewer decides how much to scrutinize, and it is designed as if scrutiny is optional. For a single agent doing a bounded task, that design works well enough. For a team of agents running in parallel, each handing artifacts to the next, the handoff screen shows one agent's output and hides the team's trajectory. The reviewer approves a node and believes they have approved the path.

The Amex note from October 13 points to one response: move the limit out of agent instructions and into the payment rail itself, so the cardholder grants a box and the rail holds it regardless of what any individual agent decides. That shifts enforcement away from the handoff screen entirely. The scaffolding cost paper from Fowler and Jain points to a different response: the harness the reviewer depends on has to be cheap enough to run and clean enough to trust, or the review step loses its foundation before the reviewer opens the screen. Both responses solve a real problem. Neither one gives the reviewer a view of the group's total in motion.

Design the team view before you design the next agent. Give the reviewer one screen that shows the group's combined spend, reach, and output, updated as the team runs, not reconstructed after it finishes. The EAIR paper's finding is that the violation was legible in aggregate and invisible in detail. Make the aggregate the default view, and let the reviewer drill into any agent's share from there.