OpenSpace: self-improving skills for Claude Code agents

My notes

Summary

OpenSpace (open-source, from HKUDS, the same team behind LightRAG and similar agent tooling) is a self-improving skill system. After a task completes, it analyses how a skill performed and decides to repair, enhance, or “capture” it as good enough. Authors claim 46% fewer tokens after the self-improvement loop runs.

Key Insight

• Skills as living artifacts, not static prompts. OpenSpace treats skills the way ML treats checkpoints: evaluate, iterate, freeze.
• Three-state lifecycle: repair (broken), enhance (working but suboptimal), captured (good enough, stop tweaking). The “captured” state is the interesting bit, since most prompt-tuning loops never know when to stop.
• 46% token reduction is the headline number. Even if optimistic, the direction matters: skills naturally bloat over time as edge cases accumulate; an automated trim/refine loop fights that bloat.
• HKUDS shipped LightRAG and similar work, so they have a track record of practical agent infra, not just papers.
• Pairs naturally with Claude Code’s skill system (.claude/skills/) where skills are markdown plus scripts that drift in quality as use cases evolve.