Agon

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:

Put the two directories side by side:

Then run Claude Code from the artifacts repository:

--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

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:

Optional local settings…