Before you publish a blueprint, make sure that your AI agent tasks are tested and producing reliable results. This section walks you through how plans work, how to test them, and how to lock them for consistent run-time behavior in Agentforce Operations.
Required Editions
| License Required |
|---|
|
This feature is a workspace that’s external to your Salesforce org and requires an Agentforce Operations license. To purchase an Agentforce Operations license, contact your Salesforce account executive. |
| Role or Access Needed | |
|---|---|
| To manage tasks in Agentforce Operations: | Admin role or Creator role |
See Also
How the AI Agents Plan Their Work
AI agents in Agentforce Operations create a plan to complete a task and reuse that plan on all subsequent workflow runs.
AI Agents Require a Plan
Tasks assigned to AI agents must have a plan to complete work. For example, an information searching task plan can include steps like searching for a supplier name in a data source, getting its information, then cross-referencing it with provided data to make sure that it matches. When a plan is locked, the agent follows the same steps in the same order on every subsequent workflow run, rather than reasoning from scratch each time. You can't publish a blueprint if it has a task that’s assigned to an AI agent without a plan.
How Plans Are Generated
There are two ways to generate a plan:
- Test tab: Open the task and run a test from the Test tab. Using the Test tab is the recommended way to generate and validate a plan before publishing.
- Run a test workflow: Run a workflow from the draft blueprint. If the task doesn't have a saved plan, the AI agent generates one during the workflow run.
Reviewing the Plan
You can review the plan in different places depending on how the plan generated. If you used the Test tab, the plan appears in the AI Agent Activity area of the Test tab. If the plan was generated during a workflow run from a draft blueprint, it's recorded in the task's Activity tab.
What Erases a Plan
Sometimes the plan doesn’t work as expected and you need the agent to use a different plan. Other times, changes you make can accidentally erase a plan. These are the ways a saved plan can be erased.
- Editing anything in the task editor.
- Deleting fields from upstream tasks or the blueprint intake that are referenced in the AI agent task.
- Editing fields through the Fields page that are referenced in the AI agent task.
- Any failure when running the task in a test workflow. For example, the agent plan expects one date format, but you provide a different format.
Test Your AI Agent Tasks
Test AI agent tasks in Agentforce Operations with realistic sample data, hone instructions, and validate plans and results before publishing blueprints.
Reliable results require testing with realistic field values and files, and then improving the instructions for the AI agent based on the results. If you test with placeholder data, such as a task named Test Task, the AI agent can give misleading results.
- In the navigation sidebar, click Blueprints.
- Click the Draft blueprints tab.
- Click Edit on the blueprint with the AI agent task to test.
- Click the AI agent task that you want to test.
- Click the Test tab.
- Click Run test and review the generated values and activity history.
-
If the task completes, but the results are incorrect, go back to the
Configure tab.
For .docx output tasks, verify that template variables exactly match the requested field names, such as
{{ supplier_name }}and{{ renewal_date }}.- Refine your instructions or other parts of the task.
- Run the test again.
-
If the task has a problem, open the task.
- Click Resolve problem.
- Select Re-run AI Agent.
- Decide whether to try the AI agent plan again or generate a new plan.
- To try again with the same plan, click Confirm.
- To try again with a new plan for testing, select Generate a new plan, and then click Confirm. You can save this plan if the results are accurate.
- If the task completes correctly, create and run a workflow with sample data from the draft blueprint to validate task interactions.
- After results are consistently accurate, save the plan for each AI agent task, then publish the blueprint to lock plans.
Locking an AI Agent's Plan
Publishing a blueprint locks the AI agent's plan in Agentforce Operations so it follows the same steps every time the workflow runs.
When you're happy with the AI agent's results and you're ready to lock the plan, publish your blueprint. The AI agent follows the locked plan every time the workflow runs.
If a locked plan fails, the agent reports a problem instead of trying a new approach. The agent reporting the problem makes failures visible and predictable so you can investigate and fix them.
If you change a blueprint, such as renaming a requested field, the saved plan can become invalid. Re-test the task before publishing again to make sure that the plan is still accurate.
To learn more about publishing a blueprint, see Publish a Blueprint in Agentforce Operations.