EvoSkill v2
A method and system for rewriting agent skill instructions at runtime through a stateful evolver, employing a three-agent architecture (baseline adapter, solver, and evolver) without modifying the base model or the runtime framework.
Timeline (last 90 days)
EvoSkill v2 is a three-agent system comprising three roles: a benchmark adapter, a solver, and an evolver.
On the SpreadsheetBench v2 benchmark, EvoSkill v2 raises held-out pass rate by 7.7 percentage points over the stronger baseline.
On the τ³-bench banking benchmark, EvoSkill v2 raises held-out pass rate by 5.4 percentage points over the stronger baseline.
EvoSkill v2's evolver is implemented as a stateful coding agent that runs throughout an entire round, equipped with bash tools to work continuously.
EvoSkill v2 optimizes skill text (instructions) rather than model weights or the runtime framework.
In the τ³-bench banking experiments, when Claude Haiku 5.5 drives both the solver and the evolver, skill gains reach as high as 21.5 percentage points.
For every percentage point of held-out pass rate improvement it achieves, EvoSkill v2's evolver costs less than the optimizers of SkillOpt and SkillBoost.
The EvoSkill v2 system accepts the skill produced by each epoch, rather than selecting among multiple candidates based on noisy scores.
EvoSkill v2's effectiveness depends on a reliable scorer and benchmark adaptation step, and the ceiling of skill evolution is limited by the quality of the scoring signal.
All Facts (9)
EvoSkill v2 is a three-agent system comprising three roles: a benchmark adapter, a solver, and an evolver.
50%UnverifiedOn the SpreadsheetBench v2 benchmark, EvoSkill v2 raises held-out pass rate by 7.7 percentage points over the stronger baseline.
50%UnverifiedOn the τ³-bench banking benchmark, EvoSkill v2 raises held-out pass rate by 5.4 percentage points over the stronger baseline.
50%UnverifiedEvoSkill v2's evolver is implemented as a stateful coding agent that runs throughout an entire round, equipped with bash tools to work continuously.
50%UnverifiedEvoSkill v2 optimizes skill text (instructions) rather than model weights or the runtime framework.
50%UnverifiedIn the τ³-bench banking experiments, when Claude Haiku 5.5 drives both the solver and the evolver, skill gains reach as high as 21.5 percentage points.
50%UnverifiedFor every percentage point of held-out pass rate improvement it achieves, EvoSkill v2's evolver costs less than the optimizers of SkillOpt and SkillBoost.
50%UnverifiedThe EvoSkill v2 system accepts the skill produced by each epoch, rather than selecting among multiple candidates based on noisy scores.
50%UnverifiedEvoSkill v2's effectiveness depends on a reliable scorer and benchmark adaptation step, and the ceiling of skill evolution is limited by the quality of the scoring signal.
50%