Multi-agent safety
The unsolved problem of keeping networks of AI agents — each acting on instructions from another — from compounding errors, deferring accountability, or being manipulated through the messages they exchange.
Learn first
When one AI agent orchestrates another, each step adds a new place for things to go wrong: a subtask is misread, a tool call returns unexpected output, or a malicious instruction injected into the environment gets passed along as if it were legitimate. The hard part is that individual agents can behave correctly in isolation while the system fails: trust, blame, and oversight become distributed across a chain that no single layer controls. No established discipline yet covers all of this — threat models, evaluation methods, and governance norms are still being worked out.
In megatrends
Finds citing this concept
Overall = breadth + depth + substance + ½·novelty + freshness (a small boost that halves each week)
A field that doesn't exist yet
DeepMind and four partners have pooled up to $10M to fund a kind of AI safety its own lead researcher says barely exists — the study of what happens when millions of agents, deployed by different actors, start interacting.
How this connects
Tap a node to open it