Agon

Claude Code plugin for autonomous AI research — from a bare topic to running experiments, with no human-written experimental code

English | 中文

Agon (paper) takes a research project from a one-line topic to running experiments. Agents plan, implement, audit and review each other in closed loops, and every handoff goes through a file on disk — so a run is recoverable, auditable, and reusable across projects. The workflow stays minimal and explicit: topic → idea → proposal → experiment.

Agon is built on Prompt Economy: treat prompt engineering as engineering, and minimize the engineering effort it demands from humans. See the paper for deployment details across more than ten research domains.

Quick start

Clone Agon and 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 --effort medium

--dangerously-skip-permissions is required because the loops are meant to run unattended: subagents write files, launch experiments, and call tools for hours with nobody at the keyboard, and a permission prompt would stall the whole run. Give Agon its own machine, container, or user account if that matters to you.

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 <topic>, and write the result to topics/mmdd-<slug>-landscape.md.
/idea-tick <topic-slug> <topic> is becoming important. Brainstorm several research ideas.
/idea-tick <idea-slug> I have a vague idea about <topic>. Create the topic file, create the idea file, and refine the idea.
/proposal-tick <idea-1> <idea-2> <idea-3> Generate proposals for these ideas.
/experiment-tick <slug> Start the experiment.
/experiment-tick <slug> 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_DEFAULT_OPUS_MODEL="deepseek-v4-pro[1m]" \
  ANTHROPIC_DEFAULT_FABLE_MODEL="deepseek-v4-pro[1m]" \
  ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-v4-flash" \
  ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4-flash" \
  CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 \
  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

The same wrapper pattern works for Codex and Grok via CLIProxyAPI (claude-codex, claude-grok).

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)"

Citation

@misc{sun2026agonautonomouslargescaleomnidisciplinary,
      title={Agon: An Autonomous Large-Scale Omnidisciplinary Research System Built on Prompt Economy},
      author={Youran Sun and Xingyu Ren and Chugang Yi and Jiaxuan Guo and Kejia Zhang and Jianda Du and Haizhao Yang},
      year={2026},
      eprint={2606.24177},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2606.24177},
}

Acknowledgements

• ARIS
• AutoResearch-SibylSystem
• Telegram MCP Communicator
• Claude Memory Manager
• humanizer
• CLIProxyAPI
• codex-switch