OpenSpace vs SkillOpt
HKUDS's skill lifecycle layer for agents: retrieve the right skill per task, evaluate which ones actually work from real outcomes, share across agents and teammates, evolve with every run. — versus — Microsoft's text-space optimizer that trains a frozen agent's skill document like weights — rollouts, bounded edits, validation-gated updates — and ships a compact best_skill.md.
Both evolve agent skills from real outcomes; openspace manages a shared skill library with retrieval, SkillOpt trains a single skill to a validation gate.
| OpenSpace | SkillOpt | |
|---|---|---|
| Stars | 7.7k | 18k |
| Forks | 922 | 1.7k |
| Language | Python | Python |
| License | MIT | MIT |
| Last activity | 2 months ago | 5 days ago |
| Topics | skills | skills, training |
| Curated connections | 4 | 4 |
OpenSpace — the curator's take
The missing half of the skill ecosystem: everyone ships installers, almost nobody closes the loop on whether a skill WORKED. OpenSpace's bet is lifecycle — retrieval at task time, evaluation from real run outcomes, and evolution instead of repeat failures — across Claude Code, Codex, OpenClaw and friends. If your skill folder has become a junk drawer, this is the category to watch. NOT yet the standard: young v2, outcome-evaluation quality depends on how honestly your runs are scored, and the Feishu/WeChat-first community signals where the early adopters are. For plain install/audit needs, a simpler skill manager does less, more predictably.
SkillOpt — the curator's take
Use it when you have a task with a scorer and want a skill that measurably improves on it — an edit lands only if the held-out score rises, and deployment adds zero model calls. Not a passive learner: you need data, a benchmark and an optimizer-model budget; the nightly SkillOpt-Sleep mode is the bridge to everyday Claude Code/Codex sessions. No metric? An in-use skill learner fits better.