Graph model flags bad agent actions before they run

A graph neural network flags a bad agent action before it executes, trained on past trajectories, per a paper posted to arXiv (2607.27443).

The payoff: an average 14.69% improvement in pass ratio across four benchmarks and three different agents, with no fine-tuning of the underlying model. That's a reliability gain teams get from data they're already collecting, not from a bigger or retrained model.

The method turns a history of agent runs into a probabilistic graph of action sequences. The GNN learns which sequences tend to precede failure. When the agent is about to repeat one, the system warns before the step runs so the agent can correct course.

If you're already logging full agent trajectories, you already have the raw material this method runs on.

The paper doesn't say what happens after the warning fires in production. It's unclear whether the agent retries automatically, whether a person approves the correction, or what latency the extra graph lookup adds per step.