Agent harness
The controlled loop Studio built so transformation analysis can run for hours, under your control.
The agent harness is the product foundation. Transformation questions — retirement impact, sequencing risk, dependency exposure — take hours across live systems. A chat box cannot run that work. The harness can.
You send a message. The agent plans, uses tools, and works on a timeline you can steer, pause, resume, and audit. Writes wait for you. A session can be reopened. A long objective can stay on a goal. A slice can be handed to a sub-agent; you stay with the original work.
What the harness does
- Planning — Draft the approach first. Keep a goal for long work. Approve, adjust, or send it back before the agent acts.
- Session — Reopen the same conversation. Compact earlier turns so the run can keep going.
- Approvals — Reads and views run. Writes arrive with a preview and wait. The agent can stop and ask you.
- Tools — Keep a checklist through the analysis. Ask you when a trade-off needs a person. Search or open a page when the question needs public context.
- Delegation — Hand an independent cut to a sub-agent. It reports back; you stay with the original work.
- Visibility — Every tool call lands on the timeline. You watch the analysis, not only the finished artifact.
Tools from connected sources (MCP) sit on Integrations. The harness tools above are part of the loop itself.
Pair with Connect your AI and Agents. The engineering contract lives in the repo at docs/HARNESS.md.