Agent definition
Configure instructions, model behavior, available tools, knowledge, output schema, and limits.
Feature guide
AI agents
Build governed agents with instructions, knowledge, tools, references, outputs, traces, and workflow participation.
On this page
Complete feature description
Understanding ai agents in UserTasks
AI Agents are governed runtime assets that combine instructions, model behavior, approved knowledge, tools, integration references, structured outputs, and operational limits. They can run interactively in a designer or page, programmatically through APIs, or as one step inside a deterministic workflow.
The agent does not receive unrestricted platform access. Designers explicitly select tools and references, while runtime authorization and tenant boundaries continue to apply. Knowledge and asset references can scope what the agent may use for a particular job.
Runs retain traces, tool calls, referenced assets, generated artifacts, output values, errors, and usage estimates. This provides the evidence required to review agent behavior and combine probabilistic reasoning with deterministic workflow controls.
When to use this feature
- The work involves classification, extraction, summarization, or natural-language interaction.
- An operator benefits from an assistant that can use approved tools.
- Unstructured content must become typed workflow outputs.
- Human review can handle uncertain or high-impact conclusions.
What you can build
- Request triage and classification
- Document extraction and summarization
- Case and knowledge assistants
- Drafting and recommendation steps
- Natural-language operational interfaces
- Agent-assisted exception investigation
Production design notes
- Keep tool access minimal and specific to the agent role.
- Define structured outputs when downstream branches depend on the result.
- Add human review for high-impact decisions and low-confidence results.
- Monitor traces, costs, failures, and unexpected tool patterns before expanding autonomy.
Core concepts
Tools and references
Allow agents to use approved integration references and platform capabilities under authorization.
Knowledge
Ground runs with selected project, document, data, or other tenant assets.
Traceability
Capture runs, steps, references, artifacts, estimated cost, and errors.
Actions and capabilities
What designers and operators can do in this product area.
AI agents actions
The main operations available in AI agents.
AI agents actions
Create and configure agent
Define its role, instructions, model, tools, knowledge, and output fields.
Example: Create a request-triage agent that returns category, urgency, and recommended workflow.
Configuration inputs
- Permission to view and edit the relevant ai agents assets.
- The target agent instructions, model, tools, knowledge, inputs, and output schema.
- Any required mappings, recipients, filters, variables, permissions, or lifecycle settings shown by the designer.
How to use it
- 1.Open Agents from the main navigation.
- 2.Create or open the relevant asset and choose “Create and configure agent”.
- 3.Complete the required fields, map dynamic values, and review access to every referenced asset.
- 4.Use preview, test, or validation where available; correct errors before publishing or executing.
- 5.Run the action and inspect its status, output, history, or audit record.
Expected results
- The governed agent can run interactively or in automation with a reviewable trace.
- Failures remain visible with enough context to correct configuration or retry safely.
AI agents actions
Run interactively
Invoke an agent from its designer or an embedded page chat experience.
Example: Create a request-triage agent that returns category, urgency, and recommended workflow.
Configuration inputs
- Permission to view and edit the relevant ai agents assets.
- The target agent instructions, model, tools, knowledge, inputs, and output schema.
- Any required mappings, recipients, filters, variables, permissions, or lifecycle settings shown by the designer.
How to use it
- 1.Open Agents from the main navigation.
- 2.Create or open the relevant asset and choose “Run interactively”.
- 3.Complete the required fields, map dynamic values, and review access to every referenced asset.
- 4.Use preview, test, or validation where available; correct errors before publishing or executing.
- 5.Run the action and inspect its status, output, history, or audit record.
Expected results
- The governed agent can run interactively or in automation with a reviewable trace.
- Failures remain visible with enough context to correct configuration or retry safely.
AI agents actions
Run from workflow
Map workflow variables into an agent and map structured outputs back out.
Example: Create a request-triage agent that returns category, urgency, and recommended workflow.
Configuration inputs
- Permission to view and edit the relevant ai agents assets.
- The target agent instructions, model, tools, knowledge, inputs, and output schema.
- Any required mappings, recipients, filters, variables, permissions, or lifecycle settings shown by the designer.
How to use it
- 1.Open Agents from the main navigation.
- 2.Create or open the relevant asset and choose “Run from workflow”.
- 3.Complete the required fields, map dynamic values, and review access to every referenced asset.
- 4.Use preview, test, or validation where available; correct errors before publishing or executing.
- 5.Run the action and inspect its status, output, history, or audit record.
Expected results
- The governed agent can run interactively or in automation with a reviewable trace.
- Failures remain visible with enough context to correct configuration or retry safely.
AI agents actions
Attach tools
Grant specific platform or integration tools instead of unrestricted access.
Example: Create a request-triage agent that returns category, urgency, and recommended workflow.
Configuration inputs
- Permission to view and edit the relevant ai agents assets.
- The target agent instructions, model, tools, knowledge, inputs, and output schema.
- Any required mappings, recipients, filters, variables, permissions, or lifecycle settings shown by the designer.
How to use it
- 1.Open Agents from the main navigation.
- 2.Create or open the relevant asset and choose “Attach tools”.
- 3.Complete the required fields, map dynamic values, and review access to every referenced asset.
- 4.Use preview, test, or validation where available; correct errors before publishing or executing.
- 5.Run the action and inspect its status, output, history, or audit record.
Expected results
- The governed agent can run interactively or in automation with a reviewable trace.
- Failures remain visible with enough context to correct configuration or retry safely.
AI agents actions
Inspect trace
Review tool calls, references, artifacts, token estimates, and execution errors.
Example: Create a request-triage agent that returns category, urgency, and recommended workflow.
Configuration inputs
- Permission to view and edit the relevant ai agents assets.
- The target agent instructions, model, tools, knowledge, inputs, and output schema.
- Any required mappings, recipients, filters, variables, permissions, or lifecycle settings shown by the designer.
How to use it
- 1.Open Agents from the main navigation.
- 2.Create or open the relevant asset and choose “Inspect trace”.
- 3.Complete the required fields, map dynamic values, and review access to every referenced asset.
- 4.Use preview, test, or validation where available; correct errors before publishing or executing.
- 5.Run the action and inspect its status, output, history, or audit record.
Expected results
- The governed agent can run interactively or in automation with a reviewable trace.
- Failures remain visible with enough context to correct configuration or retry safely.
Potential use cases
Patterns you can adapt to your own operating model.
Request triage
Classify an incoming request and recommend the right workflow path.
Form submitted
Run agent
Structured classification
Decision branch
Human review for uncertainty
Case assistant
Answer questions over approved case documents and operational data.
Embed agent chat
Scope knowledge
Grant read references
Return cited response
Preserve trace