In short
Agents built in the new Copilot Studio experience run on the Microsoft Copilot harness. Power Automate cannot start them. Use a Copilot Studio workflow (a trigger followed by an Agent node). Only agents on the standard harness are started from Power Automate with Execute Agent.
The symptom
You have an agent that triages new SharePoint items perfectly in the test pane. You build a cloud flow: When an item is created, then an action to call the agent. The Agent dropdown says No items. You try Run an agent, then Execute Agent. Both are empty. Connections are green, policies are correct.
Nothing is misconfigured. You are knocking on a door that this kind of agent does not have.
What a harness is
Every agent in Copilot Studio runs on a harness: the runtime between your design and the model. It decides when to call the model, which tools to use and what to do when a step fails. There are three.
| Harness | Built for | Started automatically by |
|---|---|---|
| Microsoft Copilot | Reasoning-heavy, multi-step business work. The default in the new Agents experience. | A Copilot Studio workflow with an Agent node, or a native agent trigger where available |
| Standard | Rule-based agents with topics and fixed paths | Power Automate, Microsoft Copilot Studio connector, Execute Agent; or event triggers on the agent |
| Customer Conversation | Website chat and phone, with human escalation | The customer, through its channel |
Before you design a trigger, find out which harness you are building on. It decides which doors exist.
The pattern that works
- Publish the agent first. Unpublished agents do not appear in the Agent node list.
- Create a workflow in Copilot Studio. Workflows, New workflow. Set the trigger type to Connector, then choose the event, for example SharePoint, When an item is created, on the right site and list. For schedules choose Recurrence.
- Add an Agent node and pick the published agent. In Message, pass the smallest thing the agent needs, usually the item ID. Let the agent fetch the rest with its own tools, so its permissions stay in one place.
- Name and publish the workflow. A name that says trigger and purpose (for example Intake, on new work item) makes the run history readable later.
- Test with a real event. Create one test item and read what the agent wrote. In our case the note appeared about 40 seconds after the item was created.
Avoiding loops
An agent that writes back to the list that triggers it can start itself again. Two simple rules prevent it:
- Trigger on created, not ** created or modified**, when the agent updates the same item. Its own update is a modification and does not fire the trigger.
- If you must react to modifications, make the agent check its own marker first (for example, a note already signed by the agent) and stop.
Then test it: create one item, wait a few minutes, and confirm there is exactly one note and one follow-up record.
What it costs
Agents and workflows on the Microsoft Copilot harness use Copilot Credits. An autonomous agent consumes credits every time its trigger fires, whether or not anyone is watching. Before you switch on a trigger, estimate events per day and check the capacity your tenant has. A recurrence every hour is 24 runs a day even when there is nothing to do, so make the agent finish quickly when it finds nothing.
Checklist
- I know which harness the agent runs on.
- The agent is published, and its tools work in the test pane.
- The workflow passes only an ID; the agent reads the rest.
- The trigger cannot fire on the agent's own writes.
- I tested with a real event and checked there were no duplicates.
- I estimated monthly runs and Copilot Credit use.
Sources
- Harnesses in Copilot Studio, Microsoft Learn
- Call Copilot Studio agents from Power Automate, Microsoft Learn
- Add an agent node to a workflow, Microsoft Learn
- Add triggers to an agent (preview), Microsoft Learn