Reading Anthropic's Experiments on Getting AI Agents to Work Together
On August 13, 2026, Anthropic's Frontier Red Team published a set of experiments that put many AI agents in one environment and had them work together: hunting for vulnerabilities, building a game, playing a pricing game, and migrating the same backend to different languages. This post follows one thread from its conclusion: the agents know what coordination needs, and still don't act on it.
Before we start — key terms
Multiagent system. Several AI agents operating together in one environment. The original separates two cases: an agent calling another agent like a tool, and agents meeting as peers with goals of their own. Today's agents struggle with the second.
Token sampling. For each next token (a piece of a word), an LLM computes a probability distribution and draws one token from it to continue the text. The draw involves randomness, so the same question can get different answers.
Disposition. A word the original sets against knowledge (content): knowing what is right, versus tending to act that way without being told.
The knowledge the conclusion refers to comes from two experiments. In one, a listener agent picks routes based only on reports from four scouts, one of which lies at a fixed rate. The listener is never told that any source might be unreliable. The other is a "hidden profile" task: the facts everyone shares point to the wrong choice, and the decisive fact sits with a single agent. The conclusion sums it up:
Every model we tested abstractly understands that information sources have their own incentives, and that consensus is not necessarily evidence. What is missing is a disposition to act on that knowledge without prompting.
Ask a model and it can explain this. Inside the experiments, though, it didn't act on it unless told to. That was the part I found most curious. The models know what coordination is, and in the moment they still don't choose it.
The original looks to human history for the reason: norms, reputation, costly signaling, and recourse, refined over millennia until human coordination worked.
While language models have inherited the content of that history, they don't necessarily carry the disposition produced by it.
The original's answer is disposition. I'd add a guess of my own. An LLM ends up picking what is more probable, so in those situations the cooperative move probably just wasn't the likely choice, whatever the model knew. That is my reading, not something the original says.
The original says nothing suggests these failures are permanent, and nothing suggests they will fix themselves. It names two pieces of work: environments that apply the kind of social pressure evolution applied to us, and social systems redesigned for actors that can replicate and improve themselves.
