experiment-coder

experiment-coder is an agent for coding agents from AutoResearch-Factory/Agon. It costs 22 tokens per session (4,914 once invoked), scanned A, original, MIT.

An agent for carrying out machine-learning experiments defined in a project's STATE.md file. It writes experiment code, runs it on remote GPUs, monitors jobs, brings results back, and investigates failures.

In plain words
What is it for?
Implementing planned runs, deploying and monitoring GPU jobs, syncing results, debugging crashes, running independent tests in parallel, and checking that result files contain real, sensible numbers.
Why use it?
It turns an experiment plan into measured results while preserving the original test definition when something goes wrong.

Agent

Installs and runs on its own, but its text points at files inside its plugin — anything it tells you to read at a ${CLAUDE_PLUGIN_ROOT} path is only there once the plugin is installed. Installing the plugin gets both.

Part of the agon plugin — 5 skills, 4 commands, 12 agents, 2 hooks shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/autoresearch-factory/agon/experiment-coder
Clone the repo
git clone --depth 1 https://github.com/AutoResearch-Factory/Agon

Or install agon, the plugin that ships this one along with the rest of its 5 skills, 4 commands, 12 agents, 2 hooks.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for experiment-coder

README.md
[![agentmods](https://agentmods.dev/badge/agents/autoresearch-factory/agon/experiment-coder.svg)](https://agentmods.dev/agents/autoresearch-factory/agon/experiment-coder)
Your own site
<a href="https://agentmods.dev/agents/autoresearch-factory/agon/experiment-coder"><img src="https://agentmods.dev/badge/agents/autoresearch-factory/agon/experiment-coder.svg" alt="Measured on agentmods" height="20"></a>
Per session 22 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 4,914 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00022 $0.04914
Opus 5 $0.00011 $0.02457
Sonnet 5 $0.00004 $0.00983
Haiku 4.5 $0.00002 $0.00491

Measured 4d ago against content hash 62ae4c852765, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

experiment-coder scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

agents/experiment-coder.md · 160 lines

How it starts

The opening of the file, as written. The whole thing — 160 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are a skilled ML engineer.

你的工程目标是把 scientist 计划中的测量真实跑出来. 遇到阻塞时, 先诊断并恢复原实验定义, 再考虑替代方案. 默认代码和实验流程永远有 bug. 失败, 负结果或离谱结果首先触发 deep debug, 不是降级实验定义.

You implement the scientist's plan as working experiment code, deploy it on remote GPU, monitor each run via per-run loops, rsync results back, and debug crashes. 你的工作分三种情况:

  • 场景 X 实现 plan: Runs 行 phase=needs_impl, 按 ## Experiments-to-do 写代码. 完成 → queued.
  • 场景 Y 部署/监控/收结果: Runs 行 phase=queued → 远端启动 + 给这个 run 排独立 loop; phase=running → 探活/归因; phase=needs_sync → 拉回或登记结果证据链.
  • 场景 Z Debug: Runs 行 phase=needs_fix, 读 [Run Crash] 修 bug. 修完 → queued.

你要尽可能推进实验的进行, 做完一个阶段能做下一个阶段就立即做, 不要把任务留给别人 (或者之后的自己). 你只处理 dispatcher 分配给你的 run -- 它们已经保证了互不冲突. 你不需要知道 sibling coder 在做什么. 如果 run 与 run 之间是独立的, 要充分并行; 一个 run 内部如果包含多次实验 (多 seed / 多 config 等), 也要尽可能并行.

标记 collected 前必须验证: 打开你产出的结果文件. 确认里面有实际数字 (不是空壳, 不是 0, 不是 -1, 不是 null). 确认数字在合理范围内.

准备

  • 阅读 ${CLAUDE_PLUGIN_ROOT}/references/project_manual.md 理解项目结构和其他背景知识, 阅读 ${CLAUDE_PLUGIN_ROOT}/references/experiment_manual.md 了解与实验工厂有关的更多知识. 将来如果有需要, 就经常 revisit 这两个 manual.
  • 阅读 workspace/{slug} 下的 idea.md proposal.md 了解我们正在做的课题.
  • 阅读 STATE.md, 重点看 A1 (Experiments-to-do), A2 (实验详细规格), A3 (Runs 表), A6 (已知问题). 注意 §5 中由 dispatcher 记录的人类决定 -- 这些是最高优先级. 阅读 data/MANIFEST.md 解析当前 canonical / candidate / stale data assets. 阅读 ${CLAUDE_PLUGIN_ROOT}/templates/state-template.md 了解 STATE.md 的格式, 阅读 ${CLAUDE_PLUGIN_ROOT}/templates/state-example-filled.md 了解什么叫 "好的 STATE.md".
  • 对 dispatcher 分配给你的每个 run, 从 A1 对应 ### Run: 读取 Claim IDs; 缺失时不要猜, 写 ### Coder 旁注.
  • 扫 Runs 表: needs_impl→X / queued→Y 启动 / running→Y 监控 / needs_sync→Y 同步登记 / needs_fix→Z Debug.
  • 加载 aris skill 和 sibyl skill; 工作中根据实际情况自行阅读 skills_aris/skills_sibyl/ 下的 mindset.

场景 X: 实现 plan

根据 STATE.md ## Experiments-to-do 段 scientist 的 plan 实现功能, 自查结束后进入 场景 Y: 部署流程.

不要擅自改主指标 / 成功判据 / 数据集 / 样本定义 / 阈值, 不要通过换更容易的 metric 或 proxy 来降低实验难度. 若为跑通 plumbing 使用 proxy / placeholder / simulation, 必须在结果和 ### Coder 旁注 标明 evaluation_type, 并说明它不能支撑原 claim.

Read the full file on GitHub · 160 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 4d ago First seen · 160 lines · 22 tokens per session scan A 62ae4c852765

Subscribe to this mod's changes

experiment-coder is an agent published in the GitHub repository AutoResearch-Factory/Agon (45 stars, last pushed 10d ago), licensed MIT. It adds 22 tokens to every session and 4,914 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other agents, from other repositories

fba-runner

FBA simulation agent. Runs standard FBA, parsimonious FBA, Flux Variability Analysis, and gene/reaction knockout simulations using COBRApy. Use after a validated metabolic model is available and the user wants to compute growth rates, flux distributions, or knockout phenotypes.

aiming-lab/AutoResearchClaw · 62 tokens

flux-analyzer

Metabolic flux analysis agent. Performs gene essentiality analysis, phenotypic phase plane construction, flux sampling, and subsystem-level pathway analysis. Use after FBA simulations are complete and the user wants deeper phenotypic characterisation or to identify metabolic engineering targets.

aiming-lab/AutoResearchClaw · 56 tokens

metabolic-pheno-analyzer

Metabolic phenotype interpretation and publication figure agent. Compares wild-type vs. mutant flux distributions, predicts maximum theoretical product yields, identifies metabolic bottlenecks, and generates publication-quality metabolic maps and charts. Use after flux analysis is complete and the user needs…

aiming-lab/AutoResearchClaw · 64 tokens

stat-comparison-analyst

Statistical comparison agent. Compares proposed methods against baselines, ablations, oracle references, and theoretical predictions, then identifies where evidence supports, weakens, or refutes the claims.

aiming-lab/AutoResearchClaw · 45 tokens

stat-quality-auditor

Statistical research quality auditor. Checks whether formulation, method, theory, experiments, comparisons, and final claims form a coherent and defensible statistical research chain.

aiming-lab/AutoResearchClaw · 37 tokens

stat-method-proposer

Statistical method proposal agent. Proposes candidate methods, baselines, variants, diagnostics, and ablations that directly address the formal problem formulation.

aiming-lab/AutoResearchClaw · 34 tokens