Building an Agent Skill Optimizer with AlphaEvolve, Harbor and Gemini 4 Argon
Google, Thursday, November 5th, 2026
Google shows how AlphaEvolve and Harbor can evolve a coding agent's skill file by scoring versions on real benchmark tasks.
Google tested a coding agent skill with OpenCode and Gemini 3.8 Flash on Terminal-Bench 2.0 tasks and found the agent passed about two-thirds of attempts with or without the skill.
To improve it, the team used AlphaEvolve, which treats the SKILL.md text as the program, has Gemini propose new versions and keeps the higher-scoring ones.
Harbor served as the evaluator, running the agent on real tasks in containers on GKE, with thousands of trials per search and notes on failures fed back to AlphaEvolve.
The approach compares each candidate against the baseline in parallel and holds out unseen tasks for a final check, yielding about 5% improvement after a few iterations.