https://github.com/AutoResearch-Factory/Agon
Autonomous Large-Scale Minimal Omnidisciplinary Research System
https://github.com/AutoResearch-Factory/Agon
Last synced: 19 days ago
JSON representation
Autonomous Large-Scale Minimal Omnidisciplinary Research System
- Host: GitHub
- URL: https://github.com/AutoResearch-Factory/Agon
- Owner: AutoResearch-Factory
- Created: 2026-06-18T14:56:15.000Z (about 2 months ago)
- Default Branch: main
- Last Pushed: 2026-07-17T05:02:09.000Z (21 days ago)
- Last Synced: 2026-07-17T07:05:27.026Z (20 days ago)
- Language: Python
- Size: 264 KB
- Stars: 25
- Watchers: 0
- Forks: 1
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- Awesome-Deep-Research - Agon - Factory/Agon?style=social) (Open-Source Implementations)
- awesome-autoresearch - Agon - Scientific research: extends Karpathy's autoresearch with Prompt Economy, organizing scientist/coder/auditor loops across more than ten domains through a modify→verify→keep/discard→repeat cycle, demonstrated in a 30-day unattended run. (Full list / Scientific Research)
- awesome-autoresearch - AutoResearch-Factory/Agon - End-to-end research orchestrator built on one cornerstone principle, Prompt Economy (reusable loops, not one-off prompts), plus five supporting rules; runs scientist/coder/auditor loops across 10+ disciplines, same reusable-loop lineage as autoresearch but scaled to full research programs.  (🔬 Research-agent systems)
- awesome-claude-skills - AutoResearch-Factory/Agon
- awesome-cli-coding-agents - Agon
- awesome-ai-agents - Agon - Factory/Agon) | Prompt Economy orchestrator: reusable scientist/coder/auditor loops instead of per-task prompts, 18 roles, 10+ disciplines | (🚀 Specialized Agents / 🔬 Research Agents)
- awesome-agent-orchestrators - Agon - Orchestrates scientist, coder, and auditor loops from research topic through proposal to experiment. (Multi-Agent Swarms)
- Awesome-Prompt-Engineering - AutoResearch-Factory/Agon - to-end research orchestrator built on one cornerstone principle, Prompt Economy (reusable loops, not one-off prompts), plus five supporting rules; runs scientist/coder/auditor loops across 10+ disciplines, same reusable-loop lineage as autoresearch but scaled to full research programs. (🔬 Autonomous Research & Self-Improving Agents / Research-Agent Systems)
README
# Agon
English | [中文](README_zh.md)
Agon ([paper](https://arxiv.org/abs/2606.24177)) is an automated research system. It keeps the workflow minimal and explicit: `topic -> idea -> proposal -> experiment`.
## Quick start
Clone [Agon](https://github.com/AutoResearch-Factory/Agon) and [agon-artifacts](https://github.com/AutoResearch-Factory/agon-artifacts):
```
git clone https://github.com/AutoResearch-Factory/Agon.git
git clone https://github.com/AutoResearch-Factory/agon-artifacts.git
```
Put the two directories side by side:
```
.
├── Agon/
└── agon-artifacts/
```
Then run Claude Code from the artifacts repository:
```
cd agon-artifacts
claude --plugin-dir ../Agon --dangerously-skip-permissions --model claude-sonnet-5[1m]
```
In Claude Code, use these commands to move the research forward:
- `/idea-tick`: create, review, refine, and literature-check ideas for a topic.
- `/proposal-tick`: turn selected ideas into reviewed proposals.
- `/experiment-tick`: coordinate scientist, coder, auditor, and reviewer roles for one workspace.
- `/deep-lit-tick`: run the shared deep literature loop used by the other stages.
## Example Prompts
```
/deep-lit-tick Exhaustively survey the literature on , and write the result to topics/mmdd--landscape.md.
/idea-tick is becoming important. Brainstorm several research ideas.
/idea-tick I have a vague idea about . Create the topic file, create the idea file, and refine the idea.
/proposal-tick Generate proposals for these ideas.
/experiment-tick Start the experiment.
/experiment-tick This is a debugging run. First explain the full procedure, then pause for my approval after each agent call.
```
## Layout
Agon itself is a Claude Code plugin. Run it from a separate data workspace, commonly named `agon-artifacts`, so prompts/code and research data can evolve independently.
Expected data workspace layout:
```
agon-artifacts/
├── topics/
├── ideas/
└── workspace/
```
Optional local settings live at `.settings.toml`. Start from `.settings.example.toml` when you need to customize model routing or parallelism.
## claude-ds installation
`claude-ds` is DeepSeek-backed Claude Code.
1. Before first launch, create an empty `~/.claude-ds/` and only handle symlinks. Claude will generate the rest on first launch:
- Settings side (symlink): `CLAUDE.md` `mcp-needs-auth-cache.json` `memory/` `plugins/` `settings.json` `settings.local.json` `skills/`
- Session side (isolated): `backups/` `cache/` `downloads/` `ide/` `stats-cache.json` `projects/` `sessions/` `session-env/` `file-history/` `history.jsonl` `paste-cache/` `shell-snapshots/` `.claude.json`
- The symlinks let `claude` and `claude-ds` share skills, MCP, and plugins
2. Add the `claude-ds()` function to `~/.bashrc`:
```
claude-ds() {
CLAUDE_CONFIG_DIR="$HOME/.claude-ds" \
ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic" \
ANTHROPIC_AUTH_TOKEN="$DEEPSEEK_API_KEY" \
ANTHROPIC_MODEL="deepseek-v4-pro[1m]" \
ANTHROPIC_SMALL_FAST_MODEL="deepseek-v4-flash" \
claude --effort max "$@"
}
```
Notes:
- Put it before the `# If not running interactively` guard, so non-interactive shells such as Claude Code's Bash tool can see it
- Add the `[1m]` suffix to the model ID to unlock 1M context
3. Run `claude-ds`
## Required Claude Code settings
| Item | Purpose |
|------|------|
| `DISABLE_TELEMETRY` | Disable Statsig telemetry (usage stats, no code or file paths) |
| `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS` | Let the main agent resume background/asynchronous subagents, and let subagents message each other |
| `cleanupPeriodDays: 3650` | Keep session history under `~/.claude/projects/` (default cleanup is after 30 days); set this at the top level of `~/.claude/settings.json` |
**Configure statusline**
Tell `claude` or `claude-ds`: "call statusline-setup, I want `[5h:6% 7d:69%(2d17h)] Ctx:7% Opus 4.6 (1M context)`"
## References
- [ARIS](https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep)
- [AutoResearch-SibylSystem](https://github.com/Sibyl-Research-Team/AutoResearch-SibylSystem)
- [Telegram MCP Communicator](https://github.com/WhymustIhaveaname/mcp-communicator-telegram)
- [Claude Memory Manager](https://github.com/WhymustIhaveaname/claude-memory-manager)
- [humanizer](https://github.com/blader/humanizer)