A scope of work is what an agent does when nobody is talking to it. One per agent, edited under Work & schedule in the editor - the first tab, with a summary of when it runs above every tab.
Objectives
One per line, concrete enough to check afterwards:
Research what competitors announced this week
Draft a LinkedIn post about the warranty and queue it for approval
Vague objectives produce vague runs. "Help with marketing" is not an objective; the line above is.
A trigger
- Only when I run it - nothing happens until you press Run now. The right default while you are still watching it.
- On a schedule - a cadence in the agent's own timezone. A missed tick is caught up on the next pass; an outage never replays every tick it slept through.
- When an event arrives - a private webhook URL. Whatever JSON you POST is handed to the agent as the event to act on. The URL is the whole credential, so treat it like a password.
Work tools
Tick only what the job needs. Research and drafting are harmless; publishing and email are the two that can reach the outside world, and they wait for you unless you say otherwise under Trust.
An agent with no tools can only write a report.
View as flow
The header of that panel has a Flow toggle. It draws the same configuration as a small diagram: what starts a run, the agent, one node per tool it may use, the approval gate when any tool can reach outside, and where the result lands.
Click any node to edit that setting - it opens the same control the form has, so the two views can never disagree. Nodes cannot be added, removed or rewired: the shape comes from the configuration, not the other way round. Every agent opens in the form view; the flow is there for when you have a few tools wired up and want the bird's-eye.
Watching the first runs
Start with Run now and read the summary the agent writes afterwards. Once two or three runs look right, give it a cadence.