Experiment with Agent Lab

Prerequisites

Have an available instrument and short historical window. Use the complete first SJS strategy, or import a saved app whose behavior you already understand. No AI connection is required for the manual baseline.

Create the experiment

Open Agent Lab and choose New experiment or Import app. Import copies a saved snapshot; it does not import running positions or keep subsequent app edits synchronized.

In Settings, select the instrument, interval, UTC start/end bounds, capital, sizing, timezone, session options, and fill delay. Paste the source in SJS and leave additional execution layers disabled.

Save and run

Choose Save revision, then Run backtest. Running also saves pending changes, but an explicit revision is a useful checkpoint. Open History, select the finished run, and inspect Chart, Trades, and Logs.

Pick the first entry attempt and explain its input, signal, request, and outcome. A finished run with no trades can be valid; check whether your source conditions actually occurred and whether the selected data initialized them.

Ask for one change

Optionally connect an authoring assistant and ask for a bounded change such as improving one diagnostic field. State which behavior should remain unchanged. Review the proposed source and its revision, then run the same window.

AI authoring assistance is separate from runtime L3 provider calls. Do not add L3 merely to use the Lab assistant.

Verify and publish

Use the reviewed run as a baseline and compare one change. Publish only the selected tested revision, then open the resulting saved app and review it.

Expected outcome: a baseline run tied to a saved revision, one explainable candidate difference, and an optional saved app. Publication does not launch a server. Common mistakes are comparing the open draft to an older run or assuming the run selected in History also selects the publication revision.

Next: compare one change.