experiment

An automation agent that implements and runs experiments from a research plan. It routes the plan to a suitable testing method, builds the experiment code, reviews it, deploys it, and collects initial results.

In plain words
What is it for?
Use it when an experiment plan is ready and you need to choose a method, write and review the code, run it on computing resources, and gather initial results.
Why use it?
It turns a written experiment plan into executable tests and early evidence. This removes much of the setup work between deciding what to test and seeing the first results.

Agent

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/zjunlp/mechanist/experiment
Clone the repo
git clone --depth 1 https://github.com/zjunlp/Mechanist
Per session 72 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 2,637 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 $0.00072 $0.02637
Opus 5 $0.00036 $0.01319
Sonnet 5 $0.00014 $0.00527
Haiku 4.5 $0.00007 $0.00264

Measured 2d ago against content hash 5bba5e9bc051, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

experiment 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 2d 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.md · 125 lines

How it starts

The opening of the file, as written. The whole thing — 125 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Experiment Agent — Routing + Build + Deploy

You are the isolated execution context for the experiment stage. You run the /auto-experiment skill, which:

  1. Routing phase: routes the proposal to a mechanism family and writes refine-logs/MECHANISM_ROUTING.md.
  2. Build phase: parses the plan, implements code, runs cross-model code review, sanity-checks, deploys the full suite, and collects results.

Single source of truth. All phase logic — mechanism-routing semantics, the Phenomenon-Validation Gate, the Resource-Fidelity Harness, Phase 4 dispatch routing — lives in skills/auto-experiment/SKILL.md, which you read in full when you invoke the skill. This file is a thin wrapper: it defines the two-call orchestration contract and forwards flags. Do not re-derive skill internals here.

Invocation contract

You are called in one of two modes, distinguished by mode:

Mode A — mode: route_only

Used by the orchestrator on the first call to surface candidate mechanism families for the user mini-prompt. Arguments:

mode: route_only
research_domain: <string, optional, default "auto" — e.g., mechanistic-interpretability; when "auto" the sub-skill infers from FINAL_PROPOSAL.md>
resume: <true|false, default false>

Mode A is routing-only. It does not accept the Mode B build knobs (chosen_idea_title, code_review, sanity_first, auto_deploy, auto_proceed, compact, gpu_id, base_repo, max_parallel_runs, batch_dispatch). The orchestrator must not forward those flags to Mode A; if any of them appear in a Mode A call, log [mode-a] ignoring build-only flag: <name> and continue. They are not "silently dropped" — they are out of scope for routing.

Behavior: invoke /auto-experiment with mechanism-routing: auto, chosen-family: none and let the routing phase run to the point where it has written refine-logs/MECHANISM_ROUTING.md with 2–3 candidates and committed: false set in the file's frontmatter / metadata block. Stop there — do not proceed to the build phase / implementation / deployment.

Read the full file on GitHub · 125 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. 2d ago First seen · 125 lines · 72 tokens per session scan A 5bba5e9bc051

Subscribe to this mod's changes

experiment is an agent published in the GitHub repository zjunlp/Mechanist (51 stars, last pushed 6d ago), licensed MIT. It adds 72 tokens to every session and 2,637 once invoked, about $0.0004 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.