autogan

简体中文

A file-based long-running agent harness for software work, built around a Generator / Evaluator loop.

Important

This project is an implementation of GAN workflow from Anthropic's article: Harness design for long-running application development.

Overview

autogan runs long coding tasks as a repeated workflow:

• a Generator proposes and implements the next step
• an Evaluator reviews it with a stricter standard
• an Orchestrator moves the process forward through files in .gan/

This helps with common long-running agent failure modes:

• drifting away from the original goal
• starting implementation before defining scope
• self-review that is too lenient
• losing context in long chat sessions

What it solves

Use autogan for tasks that:

• run long enough that a single prompt is not enough
• need an agent to review code before the work moves on
• should run as automatically as possible with minimal human intervention

It is a good fit for long-running, automated coding workflows where implementation and review both need to happen inside the loop.

Requirements

You need jq and tmux to run the workflow.

Supported agent types:

• opencode
• claude
• codex

Note

The workflow expects to run inside a Git repository and uses Git to check cleanliness and record .gan state changes.

Install

Install autogan by running:

# Install into the current directory:
./install.sh
# Install into specified directory:
./install.sh --dir /path/to/project

This creates:

.gan/
autogan.sh

Usage

1. Open a tmux session.
2. Configure your agents in .gan/config.json.
3. Make sure the git workspace is clean.
4. Start the workflow:

./autogan.sh "Build a collaborative note-taking app with comments, search, and version history."

Default config.json:

{
  "maxRounds": 10,
  "maxRepairCount": 3,
  "confirmDiscardOnRestart": true,
  "generator": {
    "command": "codex --ask-for-approval never",
    "type": "codex"
  },
  "evaluator": {
    "command": "claude --dangerously-skip-permissions",
    "type": "claude"
  }
}

Note

Context window management is delegated to the agent CLI itself. In practice, this workflow relies on the CLI's built-in auto-compaction / summarization behavior to keep long-running sessions going.

Caution

Remember to configure agent permissions in config.json. If missing, the agent may stop and wait for manual approval instead of continuing the workflow. For opencode, the configuration to skip permissions is

{
...
  "generator": {
   "command": "opencode",
   "env": {
     "OPENCODE_PERMISSION": {
       "*": "allow"
     }
   },
   "type": "opencode"
  },
...