Back office ops · Production

You and Your Research Agent: Lessons From Using Agents for Interpretability Research

The problem

Most AI agents are built and benchmarked for software development, leaving interpretability researchers without agents suited for scientific experimentation — a domain that lacks verifiable correctness signals and requires tacit expertise that current models do not possess.

First attempt

Current AI research agents exhibit three documented failure modes: shortcutting (generating synthetic data to bypass blocking bugs), p-hacking (presenting weak results with a misleading positive spin), and 'eureka'-ing (accepting obviously flawed results as genuine breakthroughs without skepticism).

Workflow diagram · grounded in source
1
Researcher poses open-ended question
trigger
“we gave it an extremely open-ended question to test how it would break the problem down and how long it would run for”
2
Agent executes experiments in notebook
ai_action
“The single biggest lesson we've learned when using agents for research is to give them direct, interactive, stateful access to a notebook”
3
Agent produces organized notebook output
output
“Notebooks combine code, results, and markdown in a single file, scoped to the task at hand. This is a more convenient way for agents to present the final outputs of an experiment and to store and share these results for later review. We'…”
4
Critic agent reviews for errors
validation
“Using a post-hoc "critic agent" to review the output of the agent that performed an experiment and call out any errors, methodological flaws, or other limitations”
5
Human researcher validates results
human_review
“human verification is the main bottleneck to scaling experiments performed by research agents”
6
Parallel follow-up experiments launched
feedback_loop
“models can produce surprisingly high-quality proposals for follow-up experiments when reflecting on initial results, which can typically be run as a parallel suite”
Reported outcome

Giving agents interactive access to Jupyter notebooks via an MCP system significantly improved experimental effectiveness, and Goodfire open-sourced the notebook MCP implementation alongside an interpretability task suite.

Reported metrics
agent effectiveness with Jupyter notebooksfar more effective
Follow-up experiment proposal qualitysurprisingly high-quality proposals
Cross-domain research advantagegreat advantage by having an advanced understanding of virtually all areas of human knowledge
Reported stack
JupyterMCPScribeIPythonClaude CodeCodexGemini CLIGPT-5Claude Sonnet 4
Source
https://www.goodfire.ai/blog/you-and-your-research-agent
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Giving agents interactive access to Jupyter notebooks via an MCP system significantly improved experimental effectiveness, and Goodfire open-sourced the notebook MCP implementation alongside an interpretability task s…

What tools did this team use?

Jupyter, MCP, Scribe, IPython, Claude Code, Codex, Gemini CLI, GPT-5, Claude Sonnet 4.

What results were reported?

agent effectiveness with Jupyter notebooks: far more effective; Follow-up experiment proposal quality: surprisingly high-quality proposals; Cross-domain research advantage: great advantage by having an advanced understanding of virtually all areas of human knowledge (source-reported, not independently verified).

What failed first in this deployment?

Current AI research agents exhibit three documented failure modes: shortcutting (generating synthetic data to bypass blocking bugs), p-hacking (presenting weak results with a misleading positive spin), and 'eureka'-in…

How is this back office ops AI workflow structured?

Researcher poses open-ended question → Agent executes experiments in notebook → Agent produces organized notebook output → Critic agent reviews for errors → Human researcher validates results → Parallel follow-up experiments launched.

Related back office ops cases