Researchers have a new way to train the scaffolding that AI agents run inside, and it starts by not trusting any single failure.
A new paper describes Ecdysis, a method for evolving the runtime harnesses that wrap large language model agents - the code that manages prompts, tools, and retries around the model itself. Existing failure-driven approaches assume any observed agent failure is evidence the harness needs a fix, but the researchers point out that a failure might instead be a quirk of that specific model rather than a flaw in the harness. Ecdysis waits, collecting failure evidence across many different tasks, and only promotes a harness change once the same failure pattern shows up repeatedly. It also runs a collaborative failure-analysis step to refine each proposed fix, spending more reasoning time during training to make modifications more precise.
That patience pays off. Across multiple LLMs and benchmarks, Ecdysis-trained harnesses were 18.56% more accurate than those built with prior harness-evolution methods, trained up to 1.84x faster, and needed less data to get there. The bigger deal is generalization - harnesses built this way transferred better across different models and used fewer tokens at inference time, rather than being tuned to whichever model happened to be running during training.
It's a reminder that in agent engineering, chasing every bug report is often worse than waiting to see which ones keep coming back.