Definition
Plain language
A method that improves a coding AI by mining its own past failures into a list of weaknesses, then manufacturing practice problems that target exactly those weak spots.
As stated in the literature
A self-evolving coding-agent framework that distills trace-derived skill cards of recurring failure patterns, uses them to inject targeted verifiable bugs into real repositories, and selects training tasks by gradient alignment with a held-out validation set rather than by difficulty.
Why it matters: It turns an agent's own mistakes into a targeted training curriculum, so practice focuses on real weak spots rather than random or merely hard problems.
For example, after noticing its coding agent repeatedly mishandles edge cases, the system writes that weakness down and plants exactly those kinds of bugs in real code for the agent to practice fixing.
Heard on the show
“The paper is called "Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills.”Episode 126 — How Coding Agents Can Mine Their Own Failures Into a Self-Targeting Curriculum