AgentDescent CLI
An agent evolution framework that edits prompts and skills with parallel workers and keeps only what passes a held-out check.
- 207 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Runs many model calls and can write modified prompts, skills and code back into the project.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/thinkflowlab-agentdescentDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Runs many model calls and can write modified prompts, skills and code back into the project.
This is third-party code. Review the repository files before installing.
What it does
AgentDescent applies the idea of gradient descent to agents themselves: the parameters are prompts, skills, agent directories, code and host plugins. N workers propose edits in parallel, one aggregator merges them into a version-controlled artifact and a held-out check decides what to keep. Edits play the role of gradients and the aggregator the role of the optimizer. It can be called from Python with a single `evolve()`, from the command line or from inside Claude Code, Codex, OpenCode, DeepSeek Harness and Grok Bot. `agentdescent demo` runs a full evolution offline without keys in about ten seconds, and `agentdescent doctor` shows which agent CLIs and keys the machine has. Twenty-three ports of published methods ship on the same engine.
Who it is for. Agent builders and researchers who improve prompts and skills on their own examples and want a held-out check.
Good fit when
- A skill fails on real examples and its rules must be improved automatically.
- Several evolution algorithms need comparing on one set of cases.
- You want to launch optimization straight from Claude Code or Codex with "improve this prompt".
- The result should be written back only after confirmation.
Not a fit when
- There is no example set or check criteria, and they must be drafted first.
- The model call budget is tight and evolution needs many runs.
- A one-time manual prompt tweak is enough.
Example request
Improve this prompt on my examples from cases.json, quote the call count in advance and write the result back only after I agree.Limitations
Python 3.9 or newer is required, and the agent-facing MCP server from the `[mcp]` extra needs Python 3.10 or higher. Research examples are outside the wheel and run from a cloned repository. Evolution spends model calls: the agent in the plugin quotes the call count before starting and waits for confirmation, and asks separately before writing back. Result quality depends on the test cases and check criteria. Grok Bot connects through a scaffold rather than in one step. The license is MIT. The last commit is dated November 2026.
How to disable. Remove the package with `pip uninstall agentdescent` and the plugins from `~/.agentdescent/plugins`. Remove the connected MCP server in the agent.
FAQ
How do I install AgentDescent?
Run `pip install "agentdescent[mcp]>=0.9.0"`. Without `[mcp]` you can work from the command line on Python 3.9.
How do I connect it to Claude Code?
Run `agentdescent install claude-code`, then `claude --plugin-dir ~/.agentdescent/plugins/claude-code`. For Codex use the `codex plugin marketplace add` and `codex plugin add` pair.
Can I try it without keys?
Yes. `agentdescent demo` runs a full evolution offline in about ten seconds: only the model is replaced and the loop is real.
What can be evolved?
Text, a skill folder, an agent folder, agent code and the host plugins themselves.
What does the agent do before launching?
Runs `doctor`, writes a spec, quotes the number of calls and waits for your consent, then confirms writing the result back separately.
Discussion
No questions or reviews yet.
Related
CLI toolsParallel self-evolution of skills and prompts: workers propose edits and an aggregator merges them into one versioned artifact
An open framework where language models and evolutionary selection find and improve code for optimization tasks
Agent self-improvement framework: optimizes prompts, memory, skills or harness code with method and evaluation set separately.
Agent memory over MCP: past work trajectories become guidelines for new tasks, with a lite version for Claude Code and Codex
cap-evolve edits an agent's prompts, tools and skills from failed traces and keeps an edit only if it wins on a held-out split
multi-agents runs a team of coding agents on a task and measures whether it beat a single agent and what it cost.
README in short
The README presents the idea and installation, then covers connecting five hosts, artifact kinds and documentation links. Measured results and algorithms live on the project site.