AI Agent
The AI Agent activity runs an artificial intelligence (AI) agent that can discover, run, and report on Cisco Workflows automations as tools. Use the activity when a large language model (LLM) needs to reason over a request and run one or more workflows to complete a task.
The AI Agent activity can run workflows autonomously. Configure the allowed workflows carefully and test the agent thoroughly before production use.
Before you begin
The AI Agent activity requires an LLM agent endpoint target and at least one workflow tool. Confirm the following before adding the activity to a workflow:
-
A target of type AI Agent Endpoint is configured and assigned to the workflow.
-
At least one model is available on the target.
-
At least one workflow or one category is selected. The agent requires either input to expose tools.
-
The workflows the agent can run are production ready.
Inputs
The activity groups inputs into four sections: Request, Agent Configuration, Tools (workflows), and Advanced Configuration. The following tables describe each field.
Request inputs
|
Field |
Required |
Description |
|
Model name |
Yes |
The model that runs the agent. The selected value always overrides the model stored on the target, so the activity runs with the model chosen here. |
|
System instructions |
Yes |
Sets the AI’s role and expertise. Shapes the tone, focus, and constraints applied to every response. |
|
User prompt |
Yes |
The question or task sent to the AI. Supports workflow variables, so the field can reference form inputs or incident data. |
Agent configuration inputs
|
Field |
Required |
Default |
Description |
|
Agent name |
Yes |
None |
Identifies the agent in the output and when auditing workflow runs. Accepts 1 to 64 characters. |
|
Maximum agent turns |
No |
30 |
The maximum number of turns the agent takes before stopping. Accepts 1 to 900 and prevents infinite loops. |
Tools (workflows) inputs
Select at least one workflow or one category so the agent has tools to run.
|
Field |
Required |
Description |
|
Specific workflows to include |
Conditional |
The individual workflows the agent can run. The list shows production-ready workflows. Required when no category is selected. |
|
Workflow categories to include |
Conditional |
The workflow categories the agent can run. The agent can run any workflow in a selected category. Required when no specific workflow is selected. |
Advanced configuration inputs
|
Field |
Required |
Default |
Description |
|
Temperature (chat response variability) |
No |
1.0 |
Controls creativity versus consistency on a scale of 0.0 to 2.0. A value of 0 returns deterministic output, and a value of 2 returns the most creative output. |
|
Max output tokens |
No |
None |
Limits the length of the AI response. Higher values increase response time and cost. |
|
Previous response ID |
No |
None |
Continues a previous agent response. Use the field for multi-turn conversations. |
|
Structured output |
No |
None |
Defines a structured output schema as a list of named fields. Each field has a name and a type of string, bool, number, or integer |
Select either specific workflows, workflow categories, or both. If neither input is set, the agent has no tools to run and the activity fails validation.
Example
The following example configures an agent that triages an alert and runs a remediation workflow.
-
In Model name, select a model from the target, such as gpt-4o.
-
In System instructions, enter You are a network triage agent. Select and run the workflow that resolves the reported alert.
-
In User prompt, reference the variable that holds the alert, such as alert.summary.
-
In Agent name, enter alert-triage-agent.
-
In Tools (workflows), select the remediation workflows the agent can run.
-
In Advanced Configuration, set Maximum agent turns to 10 to bound the run.
-
Save the activity and run the workflow.
The agent discovers the selected workflows, runs the appropriate workflow, and returns the result and run details in the output.
Outputs
The activity returns the agent response, the conversation messages, the workflows the agent ran, and token usage. The following table describes each output field.
|
Field |
Type |
Description |
|
Content |
String |
The final text response generated by the agent. |
|
Messages |
String |
The original user prompt processed by the agent. |
|
New messages |
Array |
The full set of messages exchanged during the run, including tool calls. |
|
Workflows run by agent |
Array |
The workflows the agent ran. Each entry includes a workflow run ID, a workflow name, and a status. |
|
Usage > Input tokens |
Integer |
The number of tokens in the prompt sent to the model. |
|
Usage > Output tokens |
Integer |
The number of tokens in the model response. |
|
Usage > Total tokens |
Integer |
The combined input and output token count for the request. |
Next steps
-
To send a single prompt without workflow tools, use the AI Prompt activity.
-
Reference the content output in a later activity to act on the agent response.
-
Reference the workflow_run_ids output to audit or report on the workflows the agent ran.
-
Restrict the Specific workflows to include and Workflow categories to include inputs to limit what the agent can run.