AI labs have a new way to make their agents practice on exactly the problems they're worst at.
Researchers describe EvoSyn, a multi-agent framework that generates training data for AI agents based on where those agents are actually failing. Instead of just feeding environment feedback or failed attempts back in as raw material, EvoSyn keeps a revisable map of what kinds of scenarios it knows how to generate and where it should focus next. A "Designer" role decides which gaps to target, and a "Planner" role turns that decision into concrete, executable training episodes. When the system hits a failure type it has no category for, it expands its own menu of generation options to cover it. On two agent benchmarks, tau2-Bench and CAR-Bench, agents trained this way beat the next-best synthetic-data method by 13.3 points of average task success.
That gap between noticing a weakness and knowing how to generate more of the right practice data is a real bottleneck for anyone training interactive agents, not just a theoretical one. Most prior approaches collect failures but never change the shape of what gets synthesized next, so models can plateau even as more data piles up. EvoSyn's loop (detect weakness, decide what's generatable, expand the toolkit, generate) is a more deliberate way to spend a training budget.
It's two benchmarks and a research paper, not a shipped product, so treat the 13-point jump as a promising signal rather than proof it holds up on messier, real-world agent tasks.