Five weeks of AI teaming revealed the snapshot problem in trust design

On November 2, 2026, a longitudinal study published findings that should unsettle any product team shipping an AI teammate today: workers who spent five weeks collaborating with an AI agent rewrote their handoff behavior as they went, and the trust measures taken early in that period looked nothing like the ones taken late. The study did not catch a stable attitude toward the agent. It caught a relationship forming in real time, and most of the trust signals designers rely on, short surveys, post-task ratings, A/B tests running over days, arrive before the relationship stabilizes.

The MIT ask-or-act model, published the same day, sharpens that problem from a different angle. Its finding is that the cost of an agent asking a question depends on how many other agents are running alongside it. A single clarifying question is cheap in isolation. The same question, fired by one of twelve concurrent agents, stacks onto eleven others and becomes a tax the worker pays without choosing to. Workers adjust to that tax, they stop answering, they reroute tasks, they change what they delegate and to whom. The trust relationship shifts again, and it shifts in response to the shape of the system the worker can barely see.

The meta-analysis of 420 human-AI collaboration papers adds the ceiling. Human-AI pairs outperform the average solo worker but fall short of the best solo performer. That gap is not a product failure. It marks the routing decisions that carry the most weight: the tasks where delegation loses you something real, where the best human performer would have caught what the agent missed. The five-week study and the MIT model describe how workers find that ceiling on their own, by feel, over weeks, by adjusting handoffs until the losses stop surprising them.

What workers are doing across all three of these findings is the same thing: they are calibrating delegation through accumulated experience, not through any signal the product gives them. They rewrite their handoffs because earlier handoffs taught them something the interface did not surface. They absorb the interruption tax without seeing a count. They find the ceiling without ever seeing the gap. The product stays silent while the worker does the diagnostic work the product should be doing for them.

Design the handoff surface to show the record of the relationship, not a snapshot of the last task. Surface the count of concurrent agents asking questions when a worker is deciding whether to answer or dismiss. Show the worker where their delegated tasks landed relative to the best available outcome, so the ceiling becomes visible before they spend five weeks finding it by trial. The relationship a worker builds with an agent is the design material; treat the history of that relationship as interface state, and put it somewhere the worker can read it before they hand the next task over.