Have you know about this Github Repo?

🚀 Just discovered OpenSpace, AI agents that actually learn and evolve

Came across this repo from HKUDS and it's got me thinking about the future of agent workflows.

The problem: Our AI agents (OpenClaw, Claude Code, Cursor, etc.) are powerful but they don't learn from experience. Every task starts from scratch. Same mistakes, same token burn, zero memory sharing.

OpenSpace's approach:

Skills that auto-fix when they break
Successful workflows become reusable, shareable skills
One agent learns → all agents get smarter
46% fewer tokens, 4.2x better task completion

The numbers are wild: In real-world benchmarks across 50 professional tasks (legal docs, tax returns, engineering specs), OpenSpace agents pulled $11K+ vs $2.7K using the same backbone model. Same LLM, skill memory makes the difference.

Anyone else experimenting with self-evolving agent architectures? Curious if this collective intelligence approach is the next leap or if we're still too early.

🔗 github.com/HKUDS/OpenSpace

4 comments