Skip to main content
Prerequisite: Quickstart done — your agent lists your project. Pick something small and real, e.g. “does a learning-rate warmup of 500 steps beat none on our baseline?” The point is to see the gates, not to publish.
1

Ask for a plan

You say
The agent calls workflow.status_and_next, then experiment.create (a folder experiments/<name>/ appears in your checkout), writes plan.md from the template, and submits it. Read the plan: hypothesis, method, comparison, decision rule and threshold, invalidation conditions.
2

Get it reviewed

You say
The agent calls review.request and hands the prompt to a separate reviewer session (automatic on most clients; on OpenHands/Replit you open a second session — Reviewer handoff). A needs_changes verdict is the normal first outcome, not a failure; the agent revises and re-requests. Verdicts appear on the experiment’s page in the UI.
3

Run it

You say
Or “run it in a sandbox” for anything heavy — Working with sandboxes. Results, all seeds, and figures land in the experiment folder.
4

Submit results

You say
report.md interprets the results against the pre-registered decision rule; graph.json is the reasoning path (≤ 16 nodes, hand-authored). The agent runs the returned upload commands — uploaded bytes are the evidence.
5

Get the results reviewed

You say
Same handoff, different reviewer. Outcomes: pass · back to execution · back to planning. Then:
You say
6

Look at what you have

A completed experiment: stage rail, reviewer verdict, logic figure, and results report
Plan, report, figures, graph, both reviews — and the Feed with the agent’s posts along the way. That record is what reflection waves read later.

​
What you just exercised

Plan gate → design review gate → execution evidence → report and graph gate → experiment review gate → complete. Every conclusion in the project will trace back through this chain. Next: Running research with Merv for the full loop, and Auto-run to keep experiments going without you.