Running research with Merv

You direct; the agent operates the workflow over MCP; the brain enforces the gates. State machine: Experiments and reviews.

The loop

Whatever the stage, the agent’s move is the same:

Read state

workflow.status_and_next(project_id, experiment_id?) — gates, allowed actions, missing evidence, next_action.

Do that work

Locally, in a sandbox, or through the specialist skill it names.

Submit through MCP

Plans, results, reports, graphs are real only once submitted and uploaded.

Read state again

After every transition or review.

The skills installed with the plugin teach this loop. You say what to test and answer questions.

Stage	You say	The agent does
Plan	The claim to test and the constraints (data, compute, deadline)	`experiment.create` → `experiments/<name>/`; writes `plan.md` from the template: hypothesis, method, comparison, decision rule and threshold before running, invalidation conditions. Cites prior `exp_…` ids and papers in portable form (`arXiv:…`, DOI, URL) only when they matter
Design review	”Request a design review”	`review.request` → hands the prompt to a separate reviewer (see Reviewer handoff). Rejection is normal; revise, resubmit, re-request
Execute	”Run it” — or “in a sandbox” for heavy work	Everything stays in the experiment folder. All seeds and configs kept, failed ones included. No automatic sync between checkout, sandbox, and Merv — outputs are pulled in before submission
Submit results	”Write the report and graph and submit”	`report.md` interprets results against the decision rule and stays inside the tested scope; `graph.json` is the reasoning path, ≤ 16 nodes, hand-authored. `artifact.submit` per file, every upload command run. Edited after submitting? Resubmit
Results review	”Request an experiment review”	Same handoff, `experiment-attempt-review` skill. Outcomes: pass · back to execution · back to planning
Complete	”Complete it”	Only through MCP, only when the gates hold. A rejection is followed, never worked around

Push for the smallest experiment that gives a decision-relevant signal; the plan should say why that scope is enough.

Champion: on projects with an objective, the manager agent nominates promising output with candidate.submit as soon as it appears, compares with candidate.list, promotes with candidate.promote and a reason. Files → Artifacts or Object Storage, never Git.
Reflection: after several experiments, workflow.status_and_next suggests, then requires, a wave — ask for it and the project-reflection skill runs roster → lenses → synthesis → review → publish. Thresholds: Reflection waves.
Feed: findings, kills, numbers that moved, papers, dead ends — the agent posts as it goes so the project reads without opening artifacts. The Feed.