Agents

Prepare the account and task

As a workspace owner or administrator, open AI > Models to configure an AI account and choose an available model. Model availability and usage charges depend on that account.

Open Agents in the intended business and choose Create agent. Start with a bounded task, such as summarizing an authorized ticket and identifying the information still needed.

Configure the agent

1. Write specific instructions.Describe the input, the job, and the expected result. Explain what the agent should do when information is missing.
2. Select its AI account and model.Choose an account that is available to the current business. Resolve configuration errors before publishing.
3. Choose knowledge deliberately.Enable knowledge and select ingested sources for the task. Customer-specific sources require the corresponding company context during the run.
4. Add only the tools it needs.Select operations and their verified provider connections. Decision Engine tools require a named decision engine or AI account. If you add an MCP server, review the allowed tools and mark tools read-only only when you have verified that behavior.
5. Set the output and limits.Optionally provide a JSON output schema. Set limits for model turns, tool calls, tokens, and execution seconds to suit the task.

Save, publish, and test

Save the draft, then choose Publish version. Publication checks the configuration and creates an immutable version for workflows to use. Saving later draft edits does not change that published version.

Under Run the published agent, enter controlled test input and select Start run. Provide a valid Ticket ID when the selected tools require ticket context. Use View run to inspect outputs, tool calls, errors, and any pending approvals.

Starting a test creates a real workflow execution and can incur model charges. External actions pause for approval. Review the exact action and target before approving it.

Use the published version in a workflow

Add an agent to the workflow, select the published agent, and map the input it needs. Confirm the selected version when you change the agent. Test the full workflow, including how later steps use its output.

For other workflow actions that need human review, add an explicit approval step. Instructions asking an agent to be careful do not create a workflow approval gate.

Handle a paused or failed run

• Waiting for reviewOpen the run and inspect the pending action or decision. The agent stays paused until the required review is resolved.
• Account, tool, or knowledge unavailableCheck the active business, connection verification, provider access, selected sources, and required customer context.
• A limit or output check failedReview the task and recorded output. Narrow the instructions or adjust the configured limits before starting another test.
• An external result is uncertainCheck the provider before retrying. The action may have completed even if the response did not arrive.