Hivemind open-sources a continual learning system that boosts Claude Code and Codex agent benchmarks by up to 24.8 points

@kimmonismus lets stick w the technical definition of continuous learning relating to weights and plasticity. Not md skill files. Shared context is more apt of a name imo.

@kimmonismus if u use all those agents how do u even pick which to use for a given task

@SirMrMeowmeow Fair enough

@kimmonismus the real test is whether those skills actually work across agents or if its just caching suggestions

@rugbist_ You tell me :D

@kimmonismus so all my agents finally learning from each other or just sharing the same mistakes

@kimmonismus Shared learning between AI agents feels like a behavior shift

This could mean tools that get better together and spread those improvements changing how development teams work

+19/+24 is the easy direction. The harder problem is month 3: accumulated skills start encoding last week's repo state, so without per-skill regression evals you quietly propagate stale patterns across every agent. Capture is solved; curation and deprecation are the real work. How does SkillOpt handle skill rot?

@kimmonismus Interesting approach to share learning across agents. How are you handling potential conflicts or redundant 'skills' when integrating traces from different agent types like Claude Code and Cursor?

@kimmonismus AI coding agents become better thanks to Hivemind

@kimmonismus Keep your agents secure! https://github.com/OraclesTech/guardian-sdk

@kimmonismus Best or tied on all 52 setups is what caught my eye. That consistency is rarer than the peak numbers.

@kimmonismus +24.8 in codex is wild real progress love to see it

@kimmonismus hmm curious how this handles conflicting patterns across different agents

seems like the tricky part is keeping skills useful when agent behaviors arent aligned

@kimmonismus the reusable skills push is interesting. what determines which skill gets surfaced for a given task vs ignored? that retrieval step is harder than building the skill library. http://mxlore.dev

Turning every agent run into reusable, trained skills that improve across the entire team and every tool they use is how you stop starting from zero on every new task. Most people are still treating each coding session as an isolated conversation. When traces become skills that get refined and distributed automatically, the system stops being a collection of independent agents and starts becoming a single learning organization that gets smarter with every commit.