EvoSkill v2, a new system described in an arXiv preprint, takes a three-agent approach to improving AI agents. Instead of retraining or altering the agent's model or harness, it edits the instructions—or skills—that guide the agent's behavior. The goal is to evolve these skills automatically.
A user initiates a run by describing a benchmark and supplying a grader. A benchmark adaptor then uses that input to start the adaptation process, according to the abstract. The paper's title indicates the system progresses from benchmark adaptation to a stateful evolver, suggesting a more persistent and context-aware evolution mechanism.
The work is notable for being fully agentic: the entire skill evolution pipeline is driven by agents, with no manual instruction tuning. The preprint is available on arXiv under the title 'EvoSkill v2: Fully Agentic Skill Evolution from Benchmark Adaptation to a Stateful Evolver.'