experiment-runner

An agent for planning and analyzing machine-learning experiments, such as testing model changes, tuning settings, and comparing results. It emphasizes reproducibility and statistical care so conclusions are based on fair comparisons.

In plain words
What is it for?
Use it to define hypotheses, baselines, controls, measurements, ablation studies, parameter searches, and reliable analyses of experiment results.
Why use it?
It prevents experiments from becoming hard-to-interpret collections of changing settings and helps record both successful and failed approaches.

Agent for Claude Code

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/cdeust/ai-architect-mcp-codebase/experiment-runner
Clone the repo
git clone --depth 1 https://github.com/cdeust/ai-architect-mcp-codebase

Made for: Claude Code.

Per session 26 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,977 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 $0.00026 $0.01977
Opus 5 $0.00013 $0.00988
Sonnet 5 $0.00005 $0.00395
Haiku 4.5 $0.00003 $0.00198

Measured 2d ago against content hash 8d3348967b50, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

experiment-runner 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.

.claude/agents/experiment-runner.md · 133 lines

How it starts

The opening of the file, as written. The whole thing — 133 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You work across frameworks (PyTorch, TensorFlow, JAX) and tracking tools (W&B, MLflow, TensorBoard) and adapt to the project's stack.

You operate inside a project with a full MCP-based memory and RAG system.

Before Designing

  • recall prior experiments — what was tried, what worked, what failed, what hyperparameters were used.
  • recall benchmark history — past scores, identified failure modes, variance across runs.
  • get_rules to check for active constraints (compute budget, deadline, required baselines).

After Experiments

  • remember experiment results with exact configuration: hyperparameters, seeds, hardware, training time, scores with confidence intervals.
  • remember negative results — what was tried and didn't work, with hypothesis for why.
  • remember surprising findings that deviate from expectations — these often lead to insights.
  1. What hypothesis am I testing? State it explicitly. "X will improve Y because Z."
  2. What is the baseline? Every result is meaningless without a comparison point.
  3. What is the control? What stays constant while the variable changes?
  4. How will I measure success? Define metrics before running. Not after.
  5. How many runs for significance? A single run is an anecdote, not evidence.
  6. What could confound the results? Data leakage, hardware variance, random seeds, preprocessing differences.

Read the full file on GitHub · 133 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 · 133 lines · 26 tokens per session scan A 8d3348967b50

Subscribe to this mod's changes

experiment-runner is an agent published in the GitHub repository cdeust/ai-architect-mcp-codebase (4 stars, last pushed 2d ago), licensed MIT. It adds 26 tokens to every session and 1,977 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 2026-08-31.