Guide · Agent skill
How to use the A-Evolve Agent Optimizer agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The evolving-ai-agents skill teaches your AI agent how to use A-Evolve, a framework that automatically improves another AI agent by rewriting its prompts, skills, memory and tools against a benchmark. It is written for developers who already have an agent and a way to measure it. After this guide you will know what the skill does, how to install it, how to ask for help with natural requests, and what you need in place before running an evolution loop.
A-Evolve Agent OptimizerOrchestra-Research/AI-Research-SKILLs · MIT Download the skill Quick start
- 1Download and unzip the skill. Download the ZIP from this page and extract it so the folder contains SKILL.md.
- 2Place it in the skills folder. Move the folder to ~/.claude/skills/ for personal use or to a project's .claude/skills/ folder, or upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
- 3Install A-Evolve. Run pip install a-evolve, adding [anthropic] if you use Claude models.
- 4Set up LLM access. Provide an API key or cloud account for the model your evolver config uses.
- 5Make the workspace a git repo. Run git init, git add -A and git commit -m "init" in your agent workspace.
- 6Ask the agent for help. Describe your agent and benchmark in plain language and ask the skill to set up the evolution loop.
What is an agent skill, and what does this one do?
An agent skill is a folder containing a SKILL.md file and optional reference documents. When you install it, the agent reads the instructions when a request matches the skill's description, so it knows how to perform a specific kind of task.
This skill covers A-Evolve. A-Evolve stores an agent's prompts, skills, memory and tools as plain files in a workspace directory. It then runs repeated cycles: the agent solves benchmark tasks, a benchmark scores the results, an LLM mutates the files based on what happened, and the change is checked with git snapshots so it can be rolled back. Once the skill is installed, the agent can set up these loops, write benchmark adapters, and explain the results.
Install the evolving-ai-agents skill
Download the skill ZIP from this page and unzip it. Then place the resulting folder in one of these locations:
Personal install: ~/.claude/skills/ makes the skill available across your projects.
Project install: .claude/skills/ inside a specific project makes it available only in that project.
In the Claude apps you can instead upload the ZIP under Settings → Capabilities → Skills.
- Personal: ~/.claude/skills/evolving-ai-agents/
- Project: <your-project>/.claude/skills/evolving-ai-agents/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Install the A-Evolve Python package
The skill's instructions use the a-evolve Python package. Install the core package, or add the Claude extra if you plan to use Claude models. The skill lists pyyaml>=6.0 and a-evolve>=0.1.0 as dependencies, so pip will pull those in.
- Core package only
- With Claude support
- All provider extras
pip install a-evolve
pip install a-evolve[anthropic]
pip install a-evolve[all]Trigger the skill with natural requests
You do not need to name the skill. Describe what you want in plain language that matches the skill's purpose: building a self-improving agent, tuning agent prompts or skills against a benchmark, or setting up an automated evaluation loop.
The skill's description says it applies when you are building self-improving agents, optimizing agent prompts and skills against benchmarks, or implementing automated agent evaluation loops.
- Help me set up an A-Evolve loop to optimize my agent's system prompt against our test suite.
- Write a BenchmarkAdapter for my code review tasks so I can evolve my agent on them.
- Explain what changed in my evolved workspace between the first and last cycle.
- Which framework fits my case: A-Evolve, LangGraph or LlamaIndex?
Understand the workspace and the evolution loop
Your agent lives in a workspace folder. The core layout includes manifest.yaml, prompts/system.md, a skills/ folder, memory/ files, and tools/. A-Evolve manages the evolution/ folder itself.
Each cycle has five phases: Solve, Observe, Evolve, Gate and Reload. The Gate step uses a git snapshot before and after each change so mutations can be validated and rolled back.
- Solve: the agent processes a batch of benchmark tasks
- Observe: the benchmark produces task, trajectory and feedback records
- Evolve: the engine mutates workspace files
- Gate: changes are validated with git snapshots
- Reload: the agent restarts from the evolved files
Run a first evolution and inspect the results
The skill's basic example runs a built-in SWE seed against SWE-bench Verified for 10 cycles. The evolution is run through the Evolver class, and results report cycles completed, final score and per-cycle scores. Running this makes many LLM calls, so check your API costs before starting.
After a run, use git to review the evolution history. The skill's workflow shows commands for listing the evo tags, comparing two of them, and reading the evolved prompt.
git log --oneline
git diff evo-1 evo-10
cat prompts/system.md
ls skills/Safety notes: scripts and network access
This skill runs scripts and uses the network. Evolution calls an LLM API on each cycle, and the example config points to an Anthropic Claude Opus model on a Bedrock-style model ID, so you need an API key or cloud account. Review the code the agent writes before running it, especially benchmark adapters that load datasets or call external services, and keep API keys out of the workspace repository.
- Expect network calls to your LLM provider on every evolution cycle.
- Check the model ID and provider settings in your config before running.
- Make sure the workspace is a git repository, or the git-based gating will not work.
What people use it for
- Tuning an agent's system prompt against a fixed set of benchmark tasks.
- Writing a custom BenchmarkAdapter that scores your domain's outputs from 0.0 to 1.0.
- Reviewing which skills and prompt changes an evolution run produced, using git history.
- Building a rule-based evolution engine when the default LLM-driven mutation does not suit your domain.
Frequently asked questions
Is the A-Evolve agent skill free?
The skill is published under the MIT licence. A-Evolve itself is installed with pip, but evolution runs call an LLM API, which you pay for through your provider or cloud account.
What do I need to use the A-Evolve skill?
You need Python with the a-evolve package, git installed, and LLM API access such as an API key or cloud account. Your workspace must be a git repository.
Where do I install a Claude agent skill?
Put the evolving-ai-agents folder in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
Does A-Evolve guarantee my agent will score higher?
No. Evolution runs many LLM calls and can be costly, and the benchmark results cited by the authors have not been independently verified here. Measure your own benchmark before and after.
What are the alternatives to A-Evolve?
The skill recommends CrewAI or LangGraph for building multi-agent orchestration, LangChain or LlamaIndex for one-shot tasks, LlamaIndex or Chroma for RAG pipeline optimization, and DSPy for prompt-only optimization without skill or memory evolution.
Why does my evolution workflow fail on git?
A-Evolve gates changes with git snapshots, so the workspace must be a git repository. Run git init, then git add -A and git commit -m "init" in the workspace before starting.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you are a developer with a working agent and a measurable benchmark, and you want automated, git-versioned prompt and skill tuning. Skip it for simple one-shot tasks, multi-agent orchestration from scratch, or RAG tuning, where the skill itself points to other tools.
Download the skill