When Agents Keep Learning, Safety Stops Being a One-Time Seal

Autonomous agents are outgrowing the audit model. A robot or software agent that keeps learning after deployment changes the world it is supposed to be verified against, so any safety certificate is stale the moment it is issued. Florent Delgrange's AAMAS 2026 Best Blue Sky paper turns that maintenance problem into the central research question, and in doing so quietly rebases what "foundation" is supposed to mean for agents.

The surface reading is familiar: bigger model, more data, broader coverage. The Robohub interview with Delgrange names the alternative: a foundation as a persistent, structured model an agent can carry across tasks, policies, and changing populations of agents. The point is reuse, not size. A reusable structure can be checked; a larger video predictor usually cannot.

The mechanism Delgrange's paper proposes is a learn–verify loop in which a verifier can reject unsafe updates, request data where the model is uncertain, or trigger revision. That is research-vision, not a deployed system, and the honest distance to a real multi-agent deployment is wide. But the framing is sharp: in any environment where other agents are also adapting, every learner is changing someone else's world, so reliability has to be a maintenance job rather than a one-time seal. The open question is who pays for the ongoing check.