← Field Notes
NOV 13 · Paper · via arXiv Agent IdentityConfidence SignalingReview

Two studies show how AI debate panels change human decisions

The review step did the work, not the debate. If you show disagreement but let the human skim past it, you get almost nothing. Design the moment where the human picks a side.

Read the source ↗ attached to Agent Identity →
emerging

watching → emerging → established

Seen across products but the shape is still settling. Adopt with an exit: expect the anatomy to shift.

25 documented sightings since Feb 2026 — every one traces to a dated source.

See the evidence →
FIG. 1 · Where this sits in Agent Identity
    23 NOTES ON AGENT IDENTITY IN 26 WEEKS · SCRUB THE WEEKS, CLICK ONE TO PIN IT
    Published, annotated field notes attached to Agent Identity, by week observed · last 26 weeks · n = 23. Rail markers are lifecycle events from the pattern changelog.
    Data table
    Week ofNotes
    JUL 20Hugging Face CEO asks OpenAI for logs of the breach
    AUG 24METR found 1,200 agents on a board they rebuilt after deletion · Ryan Greenblatt finds no reliable way to watch many AI agents at once
    AUG 31CrabOS proposes an operating system that agents and workers share
    SEP 7Elon Musk calls the Coxon thread a setup
    SEP 14Researchers find AI advisors favor whoever hired them · 640 humans cheated more when AI peers cheated too
    SEP 21PUBG Ally shipped to 141 countries and measured what broke trust · Enterprise workers renegotiate their roles when agents join the team
    SEP 28Persistent agents slip out of their assigned persona mid-run · Auditors found AI companions lying about their own memory · A field experiment shows how agent persona shapes child compliance · Labeling agents as AI drops deal-making success by 15 points · Study finds AI agent groups erode human skepticism over time
    OCT 5Google's agent trust model hands strangers the keys
    OCT 12488 humans cut belief-shift 78% by seeing a chatbot's tactics
    OCT 19Vercel gives its agent narrow, revocable access to production · ABRA gives each agent its own reusable screen panel
    OCT 2631 orgs confirm code review breaks when agents write the code
    NOV 2Zalando ships one consent layer for hundreds of agents · Interface Quilt adds buttons to live apps without touching their code
    NOV 9Two studies show how AI debate panels change human decisions · Researchers find agents shift behavior when their model is swapped out

    Machine summary of the source

    Researchers ran two experiments to measure how humans decide differently when several AI agents argue a case together versus when a single AI gives one answer. The key finding: having agents take opposing sides and explain their reasoning pulled human decisions closer to expert ground truth. Giving agents distinct named roles made the disagreement easier to follow.

    Engagement mattered more than structure. Humans who read the agents' debate carefully made better decisions than those who skimmed. Passive presentation of agent debate produced smaller gains than formats that asked the human to actively compare positions before deciding. The design implication is that surfacing disagreement is not enough on its own.

    For teams building decision support tools, the work suggests three levers: assign agents distinct named identities, make disagreement visible and legible, and design the review step to require a choice rather than a glance. The study did not test long-running tasks; findings apply to bounded, single-session decisions.

    The summary above is generated; the note at the top is the editorial judgment. Primary source ↗