AgentDescent plugin
Parallel self-evolution of skills and prompts: workers propose edits and an aggregator merges them into one versioned artifact
- 7 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Evolution runs call a model with your key, and token spend grows with the number of rounds and workers
- `scripts/setup-hosts.sh` changes the settings of installed agent CLIs, and the `apply` MCP tool applies the evolution result
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/agentdescentDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Evolution runs call a model with your key, and token spend grows with the number of rounds and workers
- `scripts/setup-hosts.sh` changes the settings of installed agent CLIs, and the `apply` MCP tool applies the evolution result
This is third-party code. Review the repository files before installing.
What it does
AgentDescent improves a skill, a prompt or a set of instructions using scored examples. Several workers run tasks at the same time and propose edits as diffs, while a single aggregator merges them: it resolves conflicts, fuses compatible edits, checks them statistically and commits to a Git-backed ledger. The idea is that diffs act as gradients and the aggregator acts as the optimizer. There is one entry point, the `evolve()` function of the Python library. Tasks are built from a dataset, the score is a function from 0 to 1, and the model can be any OpenAI-compatible endpoint, Claude through `claude()` or the Claude Code CLI through `claude_code()`. The result holds the learned skill itself, the final score on a held-out set and a breakdown of outcomes. As a plugin, AgentDescent connects to Claude Code, Codex and OpenCode: a shared skill tells the agent when to start an evolution, an MCP server exposes `doctor`, `plan`, `start`, `status`, `show`, `apply`, `cancel` and `resume`, and the CLI mirrors them. For a trial there is `agentdescent demo`, which runs offline with no key in about ten seconds.
Who it is for. People who maintain skills or prompts for agents and want to improve them against a measurable score instead of by feel.
Good fit when
- You have a set of examples with a check and want the skill improved so its held-out score rises
- Serial self-improvement accepts one edit per iteration and you want several workers running in parallel
- You want to start and watch an evolution from an agent: `agentdescent doctor` shows which CLIs and keys the machine has
- You want to try the idea for free: the offline `agentdescent demo` runs a full evolution without a key
Not a fit when
- There is no function to score an answer: without it the aggregator cannot accept or reject edits
- The skill is a one-off and gets edited by hand: a loop with workers and a model key will not pay off
- You need a one-step Grok Bot setup: the README describes a scaffold for it rather than automatic wiring
Example request
Run agentdescent doctor, then draft an evolution plan for my commit-message skill using the examples in data/commits.jsonl and show it without running start.Limitations
The README figures apply to specific settings: for example, 40 percent fewer model calls were measured on BBH `dyck_languages`, and the exact-match gain from 0.167 to 0.583 comes from 40 HotpotQA items with 12 held out. You will have to verify the result on your own task.
Every run calls a model with your key, and cost grows with the number of rounds and workers. Python 3.9 or newer is required. The plugin installs with `scripts/setup-hosts.sh` from a clone of the repository, and Grok Bot only gets a scaffold with an HTTPS tunnel.
How to disable. Remove the package with `pip uninstall agentdescent` and take out what `scripts/setup-hosts.sh` added: the agent skill and the MCP server block in its config. Host plugin files are created under `~/.agentdescent/plugins/`.
FAQ
How do I install AgentDescent?
The library installs with `pip install "agentdescent>=0.8.0"`, and `agentdescent demo` verifies it offline. To connect Claude Code, Codex and OpenCode, run `bash scripts/setup-hosts.sh` from a clone of the repository: the script sets up whichever agent CLIs it finds on the machine.
Which model can I use?
Any OpenAI-compatible endpoint: the address is set in `OPENAI_BASE_URL` and the key in `OPENAI_API_KEY`. The README lists DeepSeek, GLM, a local vLLM and OpenAI itself. Separate adapters exist for Claude and for the Claude Code CLI.
What does the agent get over MCP?
Eight tools: `doctor`, `plan`, `start`, `status`, `show`, `apply`, `cancel`, `resume`. An evolution started by an agent can later be inspected from a shell with the same CLI commands.
What are the gradient and the optimizer in this scheme?
The gradient is the diff a worker proposes and the optimizer is the aggregator. Unlike gradients, diffs do not add up, so the aggregator resolves conflicts and fuses edits rather than averaging them.
How is AgentDescent different from skill-creator?
skill-creator helps write a skill to a specification. AgentDescent takes an existing skill or prompt and improves it on scored examples, running edits in parallel.
Discussion
No questions or reviews yet.
Related
PluginsAn agent evolution framework that edits prompts and skills with parallel workers and keeps only what passes a held-out check.
AI Builder Club Skills
AI Builder Club — Skills
A Claude Code plugin marketplace to make a repo agent-ready and run compounding agent loops
An OpenCode plugin that runs a team of parallel agents with separate sessions, a shared task board and messaging
Anthropic skill for Claude Code that audits a multi-agent system and, by measurement, folds subagents into one loop with skills.
Runs a coding agent over tasks with known answers, studies failures and writes new skills and prompts, versioned in Git.
A tool that builds tested skills from a real workflow and governs their lifecycle through your own marketplace
README in short
The README opens with a table of measurements and names the setting of each row. Then come the diagram of workers, an aggregator and four pluggable seams, installation of 0.8.0 from PyPI, a quick start from a dataset to a learned skill and a section on connecting to agents. Confidence intervals are given where significance cannot be reached.