Devcon puts recourse after an AI agent’s mistake on the agenda

Devcon has announced a collaborative discussion with Audrey Tang and Vitalik Buterin about agent mistakes, manipulation and the limits of trusting a black-box model. The original post schedules it for November 3 at 14:00 on Lotus Stage 2, with audience questions restricted to Devcon ticket holders.

The useful product question comes before the next model upgrade: who receives a complaint when an automated decision goes wrong? A support channel without authority to inspect or stop the system offers little recourse. This is our design reading of the announcement, not a remedy Devcon says it has deployed.

For an agent allowed to move money, a reviewable permission boundary matters as much as fluent explanations. Builders should be able to describe what the agent could do, what it actually requested and which actor authorized execution, while protecting information that need not become public.

The discussion is an invitation to examine that tension, not evidence that accountability has been solved. A practical preparation exercise is to trace a plausible failure from the user’s first report to investigation and recovery. Any missing owner in that sequence is a concrete product gap.