Skip navigation
Introducing EvoSkill V1, an open-source toolkit that takes a benchmark and a coding agent, and…
0:00 / 0:36

Introducing EvoSkill V1, an open-source toolkit that takes a benchmark and a coding agent, and…

Sentient

@SentientAGI508K followers

Visit on X
161 likesDownloadOriginal post

30,234 views

Introducing EvoSkill V1, an open-source toolkit that takes a benchmark and a coding agent, and evolves it into a state-of-the-art specialist in minutes. Here’s all you need to start evolving, starting with your Claude Code ↓

56 Comments

Comments from the original post on X (Twitter)

  • Sentient's profile picture

    @SentientAGI

    1/ How does it work? Using a pattern similar to @karpathy, EvoSkill V1 acts as autoresearch for AI agent skills. It takes a benchmark, a scoring function, and a coding agent, and first evaluates the agent on that benchmark. From there, it uses the agent's failure traces as feedback to propose, test, and iterate on its prompts and skills. Read the full technical breakdown ↓

    12 likes

  • Sentient's profile picture

    @SentientAGI

    2/ Who is EvoSkill V1 for? > Devs who want to turn their agent into a task specialist without manually writing prompts > Teams building specialized solutions for customers with custom environments > Researchers working on benchmarks, agent evaluation, or autoresearch-style self-improvement

    10 likes

  • Sentient's profile picture

    @SentientAGI

    3/ Does it actually work? Yes. We ran EvoSkill on the hardest reasoning benchmarks and the results speak for themselves: > EvoSkill pushed Claude Code to SOTA performance on OfficeQA, from 60.6% to 68.1% > SealQA benchmark increased from 26.6% to 38.7%, with no human involvement, and the same skills zero-shot transferred to BrowseComp with a 5% gain > We also had similar gains with @opencode, @OpenHandsDev, @goose_oss, and @OpenAI's Codex CLI

    8 likes

  • Sentient's profile picture

    @SentientAGI

    4/ What's next? This is the first production release from Sentient Labs' AI evolution research, where we're actively exploring how to make agents self-improve across prompts, skills, memory, and the harness itself. In short, EvoSkill V1 is just the beginning. Until then, please watch & star the repo ↓

    15 likes

  • Sentient Ecosystem's profile picture

    @SentientEco

    So who else is clearing their weekend? 🙋‍♂️

    5 likes

  • FileCity's profile picture

    @FileCityAI

    Great Work! @SentientAGI Here is a visual map + narrated walkthrough of EvoSkill

    5 likes

  • JohnNguyen 🔆's profile picture

    @ThuyTrang108

    If implemented effectively, tools like these can significantly shorten the time from idea --> demo --> production.

    3 likes

  • Sentient's profile picture

    @SentientAGI

    When AI starts to build AI, so does the pace of progress!

    1 like

  • Sentient Foundation's profile picture

    @sentient_found

    self-learning at it's best!

    3 likes

  • Sentient's profile picture

    @SentientAGI

    and it's only going to get better!

  • niccomi 🧙‍♂️'s profile picture

    @0xniccomi

    sick video 🕶️

    3 likes

  • Sentient's profile picture

    @SentientAGI

    We think so too ✨

    2 likes

  • Johndyr's profile picture

    @j0hncrypt013

    cloning this tonight, gonna point it at my own lil benchmark and see what specialist pops out

    2 likes

  • Sentient's profile picture

    @SentientAGI

    Look forward to seeing it!

  • 0xMoe's profile picture

    @halaltrader_

    evolving agents against a benchmark hits different than just tuning prompts

    2 likes

  • Sentient's profile picture

    @SentientAGI

    The days of manual tuning your prompts are numbered

    1 like

  • Bryce.SOL's profile picture

    @Based__SOL

    SPECIALISTS ON TAP

    2 likes

  • Sentient's profile picture

    @SentientAGI

    It's time to let the tap flow!

    1 like

  • LuBu's profile picture

    @ZhugeLyang

    back to back bangers

    2 likes

  • Sentient's profile picture

    @SentientAGI

    we ain't done yet

    2 likes