Run and review

Once an ecosystem is compiled, operators launch runs from the web UI, watch them stream, and review the report. This guide covers that loop.

Launch a run

1. Open the frontend (terraform -chdir=infra output -raw frontend_url) and sign in.
2. Pick the ecosystem.
3. Fill in its inputs — the fields the ecosystem declared (text, select, file). Required inputs are marked; files are uploaded to the run.
4. Launch.

Watch the live run map

The run streams in real time:

• Agents light up as they start, stream output, and complete.
• Agents in the same swarm run in parallel; synthesizers wait for their depends_on agents.
• Tool calls (web search, catalog lookups, code_interpreter) appear inline as they happen.
• If the ecosystem has a human-in-the-loop gate, the run pauses for a decision and surfaces it in the attention inbox.

Read the report

When the run finishes, the report agent assembles the final deliverable. In the report / artifact viewer you can:

• Read the structured report.
• See grounding indicators — which claims are backed by source outputs (internal grounding) or web search (external grounding).
• Inspect citations — synthesizers cite upstream agents as [per <agent_id>: ...].
• Download typed artifacts an agent declared via output_format (e.g. a .md or .json file).

Give feedback

The artifact viewer supports inline section feedback: select a passage and attach a note or a score. This feedback is stored against the run and can inform iteration on the ecosystem's prompts.

Re-run and iterate

• Re-run the same ecosystem with different inputs to compare outcomes.
• If output isn't right, refine the relevant agent's prompt or dependencies (see Build an ecosystem), save to recompile, and run again.
• If agents lack grounding, populate or extend the catalog.

Where runs are stored

Run artifacts (reports, uploaded files, generated outputs) are stored in S3 and listed in the run history, along with cost and status. See Architecture for the full data and state picture.