Language agents don't slowly get worse at long tasks — they fall off a cliff.
Researchers ran a large grid search across task complexity variables: state cardinality, dependency density, horizon length, branching factor, observation mode, and mutation rate. The result is a phase diagram with three zones — a solved plateau where agents perform well, a narrow transition band, and a collapse floor where performance craters. The critical finding is mechanistic: world-state fidelity breaks down before action validity does. The agent isn't just picking bad actions; it's working from a corrupted internal picture of reality before it even decides what to do next.
This matters because the AI field has bet heavily on "agentic" systems — models that plan and act across many steps. If those systems have a hard phase transition rather than a smooth performance curve, then benchmark scores near the safe zone tell you almost nothing about behavior one step past the boundary. That's a significant problem for anyone deploying agents on real tasks with long horizons.
Stronger models push the critical boundary outward but don't eliminate the transition — which means scaling alone won't fix this. The researchers frame world-model collapse as a measurable bottleneck, which is a useful reframe: it turns a vague "agents are unreliable" complaint into something you can actually test for.