A new system called EvoSteer lets AI agent teams catch and fix their own mistakes mid-task, instead of waiting until the job is done to figure out what went wrong.
Large language model systems increasingly work in teams: an orchestrator assigns tasks to multiple AI agents and tools, wiring them into a graph of who talks to whom. Today's self-evolving versions of these systems have a timing problem. They only revise the team after a full run finishes, they spread blame for a bad outcome evenly across every step even when just one step caused it, and they add new skills to the roster without testing whether those skills are any good or retiring the ones that aren't. EvoSteer changes the lineup while a task is still running, using a scoring method called Anchored Trajectory Balance to pinpoint which specific action caused a failure, plus a try-before-you-promote process that only adds a new skill once it has passed a statistical test against a reference baseline.
That specificity is the point. An orchestrator that can identify which link in the chain broke, rather than docking the whole team equally, can fix a narrow problem without scrapping a plan that was mostly working. The researchers report gains over baseline methods across twelve benchmark datasets spanning question answering, math reasoning, code generation, and interactive decision making, which suggests the approach generalizes rather than overfitting to one kind of task.
Twelve benchmarks built and graded by the same team that built the system is not the same as surviving a messy production pipeline, where nobody is keeping score and the failures don't come in neat categories.