Create a Custom AI Worker

Create a custom worker when no template matches the job.

Start with a narrow, reviewable job. A custom worker should not be a general assistant for every process in the workspace.

Create a Worker

1. Open AI Workers.
2. Select New AI Worker.
3. Enter the worker Name.
4. Add a Description.
5. Choose the Status.
6. Write the worker Instructions.
7. Choose Context sources.
8. Choose Tool categories.
9. Confirm Output mode.
10. Set Max iterations.
11. Select Save AI Worker.

Write the name as the job the worker performs, not as a personality. Names like “Review overdue invoice follow-ups” are easier to audit than names like “Finance assistant.” The description should explain when a teammate should run the worker and what output they should expect.

Before creating the worker, write down the record type it should work on, the decision it should support, and the output a teammate will review. If those three things are unclear, the worker scope is probably too broad.

Status

• Draft keeps the worker available for setup without treating it as ready.
• Active means the worker is ready to use.
• Paused keeps the worker saved but temporarily disabled.

Use draft while writing instructions and reviewing tools. Use paused when a previously active worker should stop running temporarily.

Keep new workers in draft until they have passed at least one safe test. Move a worker back to paused before making broad changes to instructions, context, or tools if teammates may otherwise keep using the old behavior.

Output Mode

The current worker output mode is Task comment. Use workers for jobs where the output should become a structured task comment or review note.

Write instructions that fit this output. If the result should be a customer email, invoice, or workflow action, add a human review step first.

Describe the desired task comment structure in the instructions. For example, ask for a summary, evidence, risks, and recommended next step when the output is used for human review. Avoid asking the worker to make commitments to customers unless a person will rewrite and approve the message.

Max Iterations

Max iterations controls how many tool/action loops the worker can use. The form allows values from 1 to 5. Use the lowest number that still lets the worker complete its job.

Higher iteration counts can increase cost and risk. Raise the value only when a test run proves the worker needs more steps.

If the worker needs more iterations because instructions are vague, rewrite the instructions first. Iterations should support real lookup or reasoning steps, not compensate for an unclear task.

Before Activating

Run the worker on one safe task or record. Confirm context, tools, output format, and permissions before using it on real operational work.

Review the run output against the source record. Check whether it used the right context, ignored unrelated data, respected tool limits, and produced a comment a teammate can act on. If it guessed, rewrite the instructions or reduce the context before adding more iterations.

Custom Worker Checklist

Before moving a worker to active:

• the name describes the job
• instructions define what to do and what not to do
• context sources are limited to the needed records
• tool categories are no broader than the job requires
• output format is easy for a teammate to review
• one safe run has been reviewed against the source record
• the owner knows when to pause or update the worker

Troubleshooting Setup

If the worker output is too generic, add source-record requirements and an expected structure to the instructions.

If the worker references data it should not use, reduce context sources or tool categories before adding more instructions.

Related Docs

• Write Worker Instructions
• Context Sources and Tools