Evolve Where It Matters: Coverage-Guided Self-Evolution of Agent Skills
Organizations: Mohamed bin Zayed University of Artificial Intelligence · University of Liverpool
Abstract
Agent skills package procedural knowledge as natural-language instructions, and a skill helps only when its instructions are correct and properly followed by the agent. Self-evolving agents revise their own skills from the runs they make, but existing methods judge each run as a whole, by its outcome or by where it went wrong, and cannot tell whether a failure comes from a wrong instruction or from one the agent did not follow. Instruction-level coverage was introduced to test skills and labels each instruction of a run as follow, not-follow, or not covered, which separates these causes. We leverage it for skill self-evolution with SkillPlumber, a framework in which an agent evolves its own skills where coverage shows they fall short. A persistent root cause analysis agent turns each label into a revision question and checks each revision against the next run, while a maintainer distills investigations across tasks into a shared strategy library. Across three benchmarks and three executors, SkillPlumber solves 64.7% to 70.5% of the tasks, 25.6 to 30.4 percentage points higher than the initial skill and 8.7 to 14.0 percentage points higher than the strongest baseline, and removing coverage reduces recovered failures by 20.8% on SkillsBench with GPT-5.5.
Figures & tables
| Skill | Original | Mean@3 | Pass@3 | Kept | Again |
|---|---|---|---|---|---|
| Initial skill | 0 | 7.0 | 11 | n/a | n/a |
| SkillRevise | 17 | 7.3 | 12 | 41.2% | 11 |
| CoEvoSkills | 17 | 13.0 | 14 | 76.5% | 14 |
| w/o Coverage | 19 | 17.0 | 20 | 86.0% | 18 |
| SkillPlumber | 24 | 22.0 | 27 | 86.1% | 24 |
| Rate, % | Full | w/o Cov. |
| Coverage read before editing | 100.0 | – |
| R1, not-follow lines edited | 51.9 | 25.0 |
| R1, follow lines edited | 22.1 | 23.4 |
| Own previous lines rewritten | 46.4 | 37.4 |
| labeled not-follow | 56.7 | – |
| labeled follow | 40.8 | – |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting | Model | Effort |
|---|---|---|
| GPT-5.5 | gpt-5.5 | medium |
| Opus 4.8 | claude-opus-4-8 | high |
| DeepSeek-V4-Pro | deepseek-v4-pro-0813 | medium |
| Round | Ours | w/o Coverage | w/o Maintainer | Skill- Revise | CoEvo- Skills |
|---|---|---|---|---|---|
| R1 | 13 | 13 | 15 | 8 | 10 |
| R2 | 19 | 15 | 19 | 11 | 14 |
| R3 | 21 | 15 | 21 | 13 | 16 |
| R4 | 23 | 17 | 22 | 16 | 17 |
| R5 | 24 | 19 | 23 | 17 | 17 |
| SB | SLB | SWE | ||||
|---|---|---|---|---|---|---|
| Ending | Tasks | Rounds | Tasks | Rounds | Tasks | Rounds |
| Pass | 24 | 44 | 18 | 29 | 11 | 26 |
| Own stop | 8 | 30 | 7 | 22 | 2 | 8 |
| Budget spent | 20 | 100 | 4 | 20 | 23 | 115 |
| Total | 52 | 174 | 29 | 71 | 36 | 149 |
| Stopped task | Rounds |
| SkillsBench | |
| civ6-adjacency-optimizer | 2 |
| crystallographic-wyckoff-position-analysis | 4 |
| earthquake-plate-calculation | 4 |
| latex-formula-extraction | 4 |
| mario-coin-counting | 4 |
| SB | SLB | SWE | |
|---|---|---|---|
| Decisions that read a lesson | 78% | 79% | 86% |
| Decisions that cite a lesson | 65% | 73% | 98% |
| Citations from another task type | 100% | 62% | 36% |
| Citations five or more batches later | 37% | 31% | 46% |
| GPT-5.5 | Opus 4.8 | DeepSeek | |
| original execution | 35 | 34 | 30 |
| range over five | 26–35 | 27–33 | 28–33 |
| mean over five | 31.2 | 29.6 | 30.8 |
| pass@5 | 48 | 44 | 45 |
| pass@5 + original | 51 | 45 | 46 |
| passed all five | 20 | 17 | 16 |
| Method | Tokens | Recovered | Per task | |
|---|---|---|---|---|
| SB | SkillPlumber | 517.2 a | 24 | 21.5 |
| SkillRevise | n/a b | 17 | n/a | |
| CoEvoSkills | 262.1 | 17 | 15.4 | |
| SLB | SkillPlumber | 214.8 | 18 | 11.9 |
| SkillRevise | 109.3 | 15 | 7.3 | |
| CoEvoSkills | 128.7 | 8 | 16.1 |
| Tokens, M | SB | SLB | SWE |
|---|---|---|---|
| Executor | 231.3 | 104.7 | 261.1 |
| RCA agent | 196.9 | 70.6 | 176.0 |
| Coverage | 64.2 | 23.8 | 61.5 |
| Maintainer | 22.4 | 14.6 | 17.2 |
| Extraction | 2.3 | 1.0 | 2.0 |
| Total | 517.2 | 214.8 | 517.8 |
| Line label | % lines | % edits | Edit rate |
|---|---|---|---|
| All 174 edits, full method | |||
| Not-follow | 11.1 | 38.9 | 41.0% |
| Follow | 43.3 | 59.7 | 16.1% |
| No instruction | 45.2 | 0.7 | 0.2% |
| Not covered | 0.4 | 0.7 | 18.2% |
| R1, identical evidence, 52 tasks | |||
LLM Agent Skill Learning
Large language model agent skill learning studies how agents acquire, represent, reuse, and refine procedures for multi-step tasks.
See all 431 papers