EvoSkill V1: Turn Benchmarks into SOTA Agents Automatically

EvoSkill V1 is an open-source toolkit that takes a benchmark and a coding agent, then evolves the agent into a state-of-the-art specialist on the benchmark’s task in minutes.

Code Paper Technical Deep Dive

EvoSkill V1: Turn Benchmarks into SOTA Agents Automatically

Today, we are releasing EvoSkill V1, an open-source toolkit which takes a benchmark and a coding agent (like Claude Code), then evolves the agent into a state-of-the-art specialist on the benchmark’s task in minutes.

Simply run evoskill init followed by evoskill run, and your specialized AI solution or a customer demo is ready with state-of-the-art prompts and skills created from scratch by an intelligent self-improvement algorithm.

Install now into your favorite coding agent (currently supported for Anthropic's Claude Code, OpenCode, OpenHands, goose, or OpenAI's Codex CLI) and happy evolving!

What is EvoSkill?

EvoSkill is autoresearch for AI agent skills. It uses a pattern similar to Andrej Karpathy's autoresearch where EvoSkill takes a benchmark, a scoring function, and a coding agent, evaluates the agent on the benchmark, and uses the agent’s failure traces as feedback to propose, test, and iterate on the agent’s prompts and skills.

To understand the full mechanism, read our technical article here.

Results

To test, we ran EvoSkill with Claude Code and other coding harnesses on a few of the hardest reasoning benchmarks.

By installing EvoSkill into Claude Code, the agent pushed SOTA performance on:

OfficeQA from 60.6% to 68.1%
SealQA from 26.6% to 38.7%
BrowseComp — transferred a SealQA-evolved skill zero-shot for a 5.3-point gain (43.5% to 48.8%) with no human involvement

We got similarly big improvements with OpenCode, OpenHands, Goose, and Codex CLI.

Who is EvoSkill (V1) for?

If you work on building or customizing AI agents and have a benchmark, EvoSkill is for you. More specifically:

AI engineers and developers who want to turn their coding agent into a specialist on a particular task without manually writing and tuning prompts/skills themselves.
Applied AI teams building specialized solutions for a customer with a bespoke environment and success criteria.
AI researchers working on benchmarks, agent evaluation, or autoresearch-style self-improvement.
Practitioners currently using GEPA or similar prompt-optimization tools who want broader, multi-skill evolution rather than tuning one skill at a time.

Don't have a benchmark? A more general version is coming soon where we will introduce skill creation without the benchmark requirement: watch and star the repo in the meantime!

How is this different from GEPA?

GEPA was a great “first-try” at automating AI engineering. However, for automatically improving general agents like Claude Code or Codex, GEPA lacks many capabilities needed for fast iteration and effective answer quality improvement.

Namely, GEPA can't discover skills from scratch. Out of the box, GEPA can only take an existing skill, prompt, or file and evolve that in isolation. Then you would point GEPA at another skill and evolve that, and so on. On the other hand, a single evoskill run command proposes, creates, and evolves an entire multi-directory set of SOTA agent skills together, leading to significantly easier UX and output quality.

How do I use it?

You'll need three things: a supported coding agent (for now, Claude Code, OpenCode, OpenHands, Goose, or Codex CLI), a benchmark with example tasks, and a scoring function which rates the agent’s performance on the benchmark’s tasks.

Read our full installation guide and quick start walkthrough in our EvoSkill GitHub repo.

Get Involved

EvoSkill V1 is the first production tool from Sentient Labs' AI evolution research, which explores general methods to make AI self-improve — prompts, skills, memory, and even the agent harness itself.

We are always welcoming to contributions and happy to connect with other researchers and developers working on AI evolution, continual learning, and self-improvement.

Install EvoSkill and get started here: sentient.xyz/evoskill