Panel
Panel is an agentic workspace for data-science experiments.
Core trick
Four agents take turns on a Jupyter kernel, appending every plan, output review, and tag to a JSONL event log that the UI streams.
Description
Panel is an agentic workspace for data-science experiments. Instead of one AI writing cells, four specialized Claude agents — an Implementer, Interpreter, Tagger, and Archivist — run your experiment together as a sequential jury. The Implementer plans and writes code, a Jupyter kernel executes it, the Interpreter reviews the output and flags methodological risks (leakage, collinearity, temporal splits), the Tagger classifies each step against an 8-tag vocabulary, and the Archivist decides what's worth committing to a persistent knowledge base.
Every action is appended to a structured deliberation.jsonl event log. The web UI streams it live; finishing produces a shareable URL where anyone can open the experiment three weeks later and understand not just what happened, but why — the hypotheses tested, alternatives rejected, and pitfalls caught. Copilot writes cells; W&B tracks metrics; Panel captures the reasoning around them.
README excerpt
# Panel **An agentic workspace for data science experiments.** Four specialist agents — an Implementer, an Interpreter, a Tagger, and an Archivist — run your experiment together, argue over it, and leave behind a shareable artifact that explains not just *what* happened, but *why*. <p align="center"> <em>Type a goal. Watch the jury work. Get a URL you can share.</em> </p> --- ## The problem Data scientists run hundreds of experiments. Each one dies in a notebook graveyard — the code survives but the *reasoning* evaporates. Copilot writes cells; W&B and MLflow track metrics. None of them capture the deliberation: the hypothesis being tested, the alternatives considered, the pitfall that was spotted and avoided. When you open an old notebook three weeks later, you've forgotten why you pivoted from logistic regression to random forest. When you try to share your work, you paste a gist and hope. ## What Panel does differently Panel runs your experiment as a **panel of four specialized Claude agents**, one call per agent, each with its own narrow job. They run **sequentially**, each reacting to the last one's output: 1. **Implementer** writes the next notebook cell(s), records its plan, its hypothesis, and the alternatives it considered. 2. **Interpreter** reads the kernel output and explains what actually happened — flagging surprises, collinearities, leakage smells, and verification checks to try next. 3. **Tagger** assigns one or more semantic tags from a fixed 8-tag