Agent Bundle

Skill: agent-bundle

Turn the project's analysis code into an agent-callable bundle: the paper's methods become parameterized, tested MCP tools an agent or person can invoke to reproduce or extend results (Actionable), each backed by a reproduction test (Ephemerality — verifiably re-runnable). The bundle is emitted in the harness's own SKILL.md + plugin.json + MCP format, so the project dogfoods the structure it is built from. You delegate ledger/history reads and the save to the datalad doer.

Load plugins/disseminate/references/paper2agent-bundle.md for the bundle layout and mapping before generating.

When to use

• A product's analyses are provenanced (run-comparison outputs exist) and the author wants them

exposed as reusable, tested tools / an MCP server.

• Do NOT use to write the paper (draft-manuscript) or to build the reproducible article

(executable-article) — this exposes the methods as callable tools.

Steps

1. Extract candidate tools — from code/ scripts and the participants.json data dictionary:

each analysis becomes a parameterized tool (inputs = the script's real inputs, typed from the dictionary; outputs = the provenanced derivatives/…). List them and confirm scope with the user.

1. Scaffold the bundle (per the reference) at agent-bundle/: .claude-plugin/plugin.json,

one skills//SKILL.md per tool, mcp/server.py exposing the tools over MCP, .mcp.json registration, and tests/reproduce.py.

1. Write reproduction tests — each test asserts the tool reproduces the recorded result for its

comparison (derivatives/cmp-/), so the bundle is verifiably re-executable.

1. Register + log — add the bundle path to the product's outputs[] (or create an

agent-bundle-kind product); append { ts, op: agent-bundle, stage: disseminate, note: "Paper2Agent bundle for ", branch: }.

1. Save — delegate to the datalad doer: "save: datalad save -m 'agent-bundle: synthesize '."
2. Report — the bundle path, the tools exposed, how to run the MCP server + reproduction tests,

and the next step: link-outputs to relate the bundle to the code (IsDerivedFrom) and paper (IsSupplementTo).

Constraints

• Tools wrap existing provenanced analyses — do not invent new methods or change the science; a

tool's output must match the recorded result (that is what the reproduction test asserts).

• Emit the harness's own SKILL.md + plugin.json + .mcp.json format (dogfooding) — do not

invent a bespoke bundle format.

• Every tool ships a reproduction test; a bundle whose tools are not verifiably reproducible is not

releasable.

• Record the bundle under the product's outputs[]; keep log: append-only and the ledger

schema-valid. Delegate reads/saves to the datalad doer.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: neuroanalytics
• Source: neuroanalytics/data-science-harness
• License: MIT

Install and usage instructions live in the source repository linked above.