Skip to content

Agon

external · status: active · focus: end-to-end · discipline: general · started: 2026

Project page: https://github.com/AutoResearch-Factory/Agon

Licence: MIT

Source: projects/landscape/agon.yml

Positioning

A Claude Code plugin that drives a research project from a one-line topic through idea → proposal → running experiments with no human-written experimental code. Twelve role-scoped subagents (idea creator/refiner/reviewer, proposal refiner/reviewer, experiment scientist/coder/auditor/reviewer/screener, deep-lit reader, env-validator) hand off work exclusively through files on disk in a separate "artifacts" workspace, so a run is recoverable, auditable, and resumable across sessions. Sits at the literature → ideation → research-design → code/analysis slice of the pipeline; the project's own rules explicitly forbid the scientist role from ever drafting the paper, so paper-writing and dissemination are out of scope for the public plugin.

Distinctive contribution

"Prompt Economy" (arXiv:2606.08878) applied concretely: rather than one big autonomous loop, Agon decomposes research into tick-based commands (/idea-tick, /proposal-tick, /experiment-tick, /deep-lit-tick) run unattended for hours via --dangerously-skip-permissions, with adversarial auditor/reviewer subagents that re-read raw result files (not the author's summarized claims) to catch drift, downgraded claims, or numbers that don't match logs before a round can be marked reviewed. Built on top of ARIS's skill and the AutoResearch- SibylSystem (both credited, not reimplemented), and reports deployment across 10+ research domains in its arXiv paper (2606.24177).

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Six stages (literature discovery/synthesis, hypothesis generation, research design, code generation, data analysis); explicitly excludes paper-drafting and dissemination by design ('scientist agent never writes the paper').
Autonomy level 3 README states loops 'run unattended... for hours with nobody at the keyboard'; requires --dangerously-skip-permissions and recommends giving Agon its own machine/container.
Architectural transparency 3 All 12 agent prompts, 5 skills, hooks, and templates are public Markdown/Python in the repo; arXiv paper documents the Prompt Economy design principles.
Inputs supported 2 Single input form (a topic or vague idea in text) but multi-source literature access (arXiv, Semantic Scholar, OpenAlex, Google Scholar via arxiv-tools skill); no external structured-data-source integration observed.
Outputs / reproducibility 2 Every handoff persists to disk (idea/proposal/STATE/experiment-log files, code, results) and is explicitly designed to be resumable/auditable, but there is no packaged reproducibility manifest and no final paper artifact.
Internal evaluation 1 Paper describes narrative deployment across 10+ research domains; no reported benchmark scores, leaderboard, or third-party replication in the repo or abstract.
Openness 2 MIT license, public repo, runnable by anyone with a Claude Code (or DeepSeek/Grok/Codex-proxied) subscription; not fully commodity-hardware-free since it depends on a paid LLM backend.
Maturity / traction 1 Young (paper June 2026, repo active through Aug 2026), ~44 stars / 4 forks at time of review; single-team research prototype, pre-1.0.
Cross-family policy 1 Optional — README documents a claude-ds wrapper (DeepSeek) and notes the same pattern works for Codex and Grok via CLIProxyAPI, but the default path is Claude models throughout.
Runtime assurance 2 Multiple in-pipeline gates: experiment-auditor checks result files for sanity/timestamp/contradiction before claims stand, experiment-reviewer independently re-reads code/logs against claims with full repo access, and a separate novelty-check skill runs multi-source literature verification on proposals.
Cross-platform portability 2 Ships as a Claude Code plugin, but the documented backend-swap pattern (claude-ds, claude-codex, claude-grok via CLIProxyAPI) gives it 3+ effective LLM providers; still a single agent runtime (Claude Code CLI).

Scored on 2026-09-01. See the evaluation rubric.

Tags

Pipeline stages: literature-discovery literature-synthesis hypothesis-generation research-design code-generation data-analysis

Architectural features: multi-agent tool-use iterative-loop persistent-memory artifact-versioning

Inputs: one-line-topic vague-idea-description

Outputs: idea-files proposal-files experiment-code experiment-logs-and-results

Data sources: user-provided

Knowledge sources: arxiv semantic-scholar openalex google-scholar

Limitations

  • No paper-writing or dissemination stage in the public plugin — the design explicitly forbids the scientist role from drafting manuscripts, so output is code + results + logs, not a submittable paper.
  • Heavily Chinese-language agent prompts and internal docs; English-only users get the README/commands but not the full prompt text without translation.
  • No reported benchmark evaluation — claims of effectiveness rest on narrative deployment description in the arXiv paper, not a scored comparison.
  • Requires --dangerously-skip-permissions and hours of unattended tool/code execution; safe only in an isolated machine/container as the README itself warns.

Papers describing this project

  • Agon: An Autonomous Large-Scale Omnidisciplinary Research System Built on Prompt Economy — Sun, Youran, Ren, Xingyu, Yi, Chugang, Guo, Jiaxuan, Zhang, Kejia, Du, Jianda, Yang, Haizhao (2026). arXiv preprint. arXiv:2606.24177