Microsoft Research releases SkillOpt, an optimization method that treats AI agent skills as trainable external states of a frozen model

These concepts coming soon to GBrain this week

can't wait for this gbrain feature. here's the loop: agent attempts a task using a skill ↓ gbrain eval or LLM-as-judge scores the result ↓ dream cycle runs the optimizer overnight ↓ proposes small edits to the SKILL.md ↓ if the new version scores higher, accept ↓ commit the improved skill, next run uses it

@omarsar0 Hell yeah this is awesome

@omarsar0 How I can eval my skills ? I should i rerun same prompts with loaded rewritten skill on git checkpoint ? How u guys automate that to get real evals? I appreciate any good blog, paper, repo on that topic

I don't have this paper in my Arxiv archive yet but I matched on it and related topics, notes, questions, and papers are here if you and others are interested. I ported my Obsidian vault of ~1400 paper excerpts online & added cross-cutting connections to help w research/discovery. https://whitepapers.gravity7.com/match/?arxiv=2605.23904

@omarsar0 Skills transferring across models and harnesses is the real punchline here. The compounding mechanism isn't better weights — it's the tool layer learning to describe itself.

@omarsar0 You might be interested in the stuff I am doing over at http://oagp.org

Basically, designing a structured, inheritable substrate for agentic governance. It's actually pretty miraculous when you use it.

@Jaimieborodo working on a write-up to answer this exactly. stay tuned!

@garrytan this sounds like youre building an actual brain instead of just a bot script

curious how the feedback loop looks in practice

@Voxyz_ai That's gr8 but bro I think I will go broke with all these agents 😭.... Fking 200+ USD every month on ai is too much now and I noticed only 5-10$ of those are productive rest is bs...

@garrytan Ok I’ll bite, @grok what is gbrain 👀

GBrain is garrytan's open-source AI agent "brain" (http://github.com/garrytan/gbrain).

It's a Postgres + pgvector memory layer that gives agents like OpenClaw/Hermes perfect recall + synthesis over 10k+ markdown files (notes, meetings, emails, etc.). It auto-builds a typed knowledge graph, does hybrid search, and runs cron jobs autonomously.

Garry just teased SkillOpt-style skill optimization coming to it this week. Super useful for production agents.

@satyaa_goyal lmao rip. i tell myself it's r&d

@garrytan Same finding on my end - the self-improving loop isn't just the skill docs, the whole knowledge layer compounds: data, hypotheses, rules, procedures. An example from march, 2026 - simplified.

Actually, I often replace skills with custom files to control what loads when better.

@omarsar0 lol I created this. But called it skillopts

Right now I just use my personal AI and our company brain and it screws up and I tell it to fix it and write tests for it.

Also I do cross modal evals on progressive batches (eg if there are 10000 items do 5 and eval the input and output and skill, then keep doubling the batch size as you go)

@omarsar0 @grok làm rõ vấn đề hơn cách đơ giản

@omarsar0 can’t wait for the optimizer to decide that the best "skill" is just to tell the user to do it themselves. peak efficiency

@omarsar0 Interesting direction-treating the skill document as an optimizable external state feels much more scalable than manually crafting prompts. The cross-model transfer results are especially compelling and worth deeper exploration.