We taught Hivemind skills to get better on their own

When we launched Hivemind, the chain was simple: capture what your coding agents do, codify the repeated patterns into reusable skills, and propagate those skills across every agent your team runs.

It worked. But it was missing a step. Skills accumulated, and a growing pile of skills is not the same as skills that get better.

Today we're closing that gap. SkillOpt is now built into Hivemind.

SkillOpt is a text-space optimizer out of Microsoft and collaborators. Instead of just saving a skill, Hivemind now trains it: it scores the sessions where a skill actually got used, keeps the changes that help, and drops the ones that don't. The skill gets sharper over time instead of bloating. The model itself never changes. This is continual learning at the skill layer.

In the paper, that approach added +19.1 points of accuracy inside Claude Code and +24.8 inside Codex, and won or tied on all 52 setups tested.

Everything else you already know still holds. It works across Claude Code, Codex, Cursor, OpenClaw, Hermes, and pi. And it all runs on your cloud. Your traces and your skills stay in your own cloud storage.

If you're already running Hivemind, update and SkillOpt is there. If you're not yet:

npm install -g @deeplake/hivemind && hivemind install

Try it:

https://github.com/activeloopai/hivemind 

The SkillOpt paper: arxiv.org/pdf/2605.23904

As always, reply and tell us what's working and what isn't. We read everything.

Regards,

Davit