SkillHone: A Harness for Continual Agent Skill Evolution Through Persistent Decision History

2026-06-07 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

SkillHone is a novel harness designed for the continual evolution of language-model agent skills, addressing the limitation of existing methods that discard crucial decision history. It facilitates cross-session refinement by pairing skill revisions with evaluation-side evidence, recording structured histories of diagnoses, revisions, evidence, and outcomes. This system employs role-separated subagents to test candidate skills on practice probes, proposing revisions informed by past decisions and eliminating the need to rediscover prior rationale. Evaluated on deep-research benchmarks in a raw open-web setting, SkillHone, utilizing Qwen3.6-35B-A3B as its backbone, significantly outperforms a deep-research agent backed by commercial retrieval services. It achieved a 15.8-point improvement on GAIA and a 3.2-point improvement on WebWalkerQA-EN, also surpassing previous skill-evolution techniques.

Key takeaway

For AI Engineers developing continually evolving agents, SkillHone offers a robust framework to overcome the challenge of lost decision history. You should consider implementing persistent decision logging and structured feedback mechanisms to enable agents to learn from past revisions and evaluations. This approach prevents redundant rationale discovery, significantly improving agent performance on complex, dynamic tasks like open-web research.

Key insights

SkillHone enables continuous agent skill evolution by persistently recording and utilizing decision history and evaluation feedback.

Principles

• Preserve decision history for agent learning.
• Integrate evaluation feedback directly into revisions.
• Use role-separated subagents for refinement.

Method

SkillHone records structured histories of diagnoses, revisions, evidence, and outcomes. Role-separated subagents run candidate skills on practice probes, proposing revisions informed by prior decisions for cross-session refinement.

In practice

• Implement persistent decision logging for agents.
• Design subagents for iterative skill refinement.
• Integrate evaluation feedback into agent training loops.

Topics

• Agent Skill Evolution
• Language Model Agents
• Decision History
• Continual Learning
• Open-Web Research
• GAIA Benchmark
• WebWalkerQA-EN

Best for: Research Scientist, AI Scientist, AI Engineer, Machine Learning Engineer

Related on AIssential

• SkillClaw: Let Skills Evolve Collectively with Agentic Evolver — SkillClaw enables LLM agents to collectively evolve skills by aggregating multi-user interaction data.
• Learning to Choose: An Empowerment-Guided Multi-Agent System with semantic communication for Adaptive Method Selection — Reliable autonomous learning in multi-agent systems requires preserving action-outcome fidelity through semantic checkpoints to prevent drift.
• How to Make Claude Code Improve from Its Own Mistakes — Enabling AI agents to continually learn from their mistakes is a key challenge for improving their long-term performance.
• New framework lets AI agents rewrite their own skills without retraining the underlying model — Memento-Skills enables AI agents to autonomously evolve their capabilities by rewriting their own executable skills.
• RECAP: Regression Evaluation for Continual Adaptation of Prompts — Existing prompt adaptation methods are structurally inadequate for proactive, real-time constraint evolution in agentic systems.
• Skills on the Fly: Test-Time Adaptive Skill Synthesis for LLM Agents — Dynamically synthesizing task-specific skills from retrieved trajectories improves LLM agent performance without parameter updates.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.