Create your first agent

Start with one job

Open Agents and describe the job you want to delegate. Include the input, expected output, tools and limits. An agent needs instructions it can act on, not just a title such as ‘marketing assistant’.

Configure and test

Create an agent that reviews public competitor announcements and prepares a weekly summary with sources. It should flag uncertain claims and never publish or contact anyone. I will review the first run.

1. Review the proposed name, instructions, tools and permission setting.
2. Connect any required private service and select only the tools relevant to the job.
3. Save the agent and open its configuration to confirm the recorded instructions.
4. Request one execution and inspect the result and any pending approvals.
5. Adjust the instructions before adding a recurring routine.

Change an existing agent

Name the existing agent when requesting a change. Say which instruction, language or tool should change and what should be preserved. Adding a requirement should update the agent rather than create a second one.

An agent’s autonomy setting does not override the permissions of its tools. A task may pause for approval, unavailable access or a credit limit. Check the execution history for the actual outcome.

Check the saved job description

Open the agent’s sheet after creation. Read the job description and when the agent should be called in. Confirm the output language, allowed sources, intended destination and how it should handle missing information. These are the instructions for later executions, so a requirement mentioned only in an unrelated conversation should not be assumed to be saved.

For an existing agent, request a focused update: “Add French as the response language for the weekly analyst. Keep its sources, reporting format and permissions.” Then reopen its sheet to check the change. Editing an agent and launching an execution are separate actions.

Run a test with a clear acceptance check

1. Ask for one execution of the named agent with a small, identifiable input.
2. Use execution history to distinguish a queued run, an active run and a request awaiting approval.
3. Compare the result with the job description: sources, language, format and destination.
4. If the test fails, correct the specific instruction, access or permission before repeating it.
5. Add a routine only after this manual result works. Review its cadence and the account’s available credits.

Next steps

• Instructions, skills and memory
• Permissions and approvals
• Routines and execution history