Deploying an Agent in a Workflow
An agent isn't deployed on its own — it's attached to an AI Agent activity on a workflow canvas, and it's the workflow that gets deployed and exposed to whatever calls it.
Before you begin
Publish the agent first. Only published agents are offered when attaching one to a workflow, since an open draft's connection and other references aren't safe to embed while they can still change under it. See Publishing an Agent.
Attach a published agent
- Add an AI Agent activity to a workflow. See AI Activities.
- In the activity's configuration, select an agent from the field that selects an agent.
Selecting an agent copies in a frozen snapshot of its published configuration — connection, model, instructions, sampling parameters, memory settings, knowledge sources, tools, and guardrails — so later changes to the agent's draft don't affect a workflow already using it. Publish again and reattach the agent to pick up changes.
How you know it worked
The activity's fields fill in from the selected agent's published configuration.
When it does not work
| Symptom | Cause | What to do |
|---|---|---|
| Failed to load agents | The list of agents couldn't be fetched | Try again |
| Could not fetch this connection's API key — enter it manually below | The agent's connection secret couldn't be retrieved automatically | Enter the API key manually |
| This agent has no connection configured yet — open it in the Agent Builder first | The selected agent has no connection attached | Attach a connection from the agent's Capabilities tab, then publish it |
Next steps
- Reviewing Logs — every reply the agent gives, whether from a test chat or a deployed workflow, shows up here.
- Configuring Memory — inspect what a deployed agent has actually written to memory.