agentagon

Agentagon — AI skill for Claude Code

AI community

Evidence-backed audits, evaluations, and measured improvements for any AI agent.

How to install Agentagon

This entry records only its repository, not the path inside it, so there is no exact command to give. Open agentagon/agentagon and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Agentagon does

Evidence-backed audits, evaluations, and measured improvements for any AI agent.

Alternatives in AI

Nodeterm — Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as dragga 1.3k ★
SEO Geo Claude Skills — 20 SEO & GEO skills for Claude Code, Cursor, Codex, and 35+ AI agents 888 ★
Academic Writing Skills — AI-powered post-writing toolkit for academic papers — format validation, grammar/style polishing, de-AI editin 430 ★

README

Agentagon

**Recursive self-improvement for AI agents.**

Improve agents using code, traces and verified experiments. Remember successes, failures and production outcomes to guide the next attempt.

Agentagon has five concepts: a **dashboard** for people, a managed coding **brain** for reasoning, **memory** groups for lessons, built-in **workflows** for repeatable work, and shared **capabilities** for evidence and execution. Dashboard and MCP use one local service and one task runtime.

flowchart LR
    Dashboard --> Runtime[Workflow runtime]
    MCP --> Runtime
    Runtime <--> Brain[Managed Codex or Claude]
    Runtime <--> Memory[Local memory groups]
    Runtime --> Capabilities[Traces, evaluations, execution and delivery]

Install and start

Requires macOS or Linux, Python 3.12+, Git for executable evaluations, and an authenticated Codex CLI or the optional Claude Agent SDK with API-key access.

pipx install agentagon
cd /path/to/your-agent
agentagon

`agentagon` opens the React dashboard and starts or reuses a detached local service. Closing the browser or terminal does not cancel tasks. `agentagon serve` runs that service in the foreground; `agentagon mcp` serves stdio MCP without opening a browser. Optional extras: `agentagon[claude,credentials]`.

Set up your improvement loop

Open any existing local folder, or choose the separate clone action when you need a new checkout. A Git repository or GitHub remote is not required for assessment. **Detect agents** scans the folder and makes detected agents available immediately, without a model or an acceptance queue. Open an agent, choose a goal, optionally add details, and click **Go**. Agentagon coordinates measurement and improvement in the background and asks for missing inputs. Code and trace assessment remains an optional deeper action.

The persistent workspace is organized around **Overview**, **Agents**, **Production**, and **Activity**, with project configuration under **Se