Start with one verifiable task
A useful task has an expected output and a destination. Ask for a comparison with public sources, a draft based on an authorized document or an update in a connected tool.
The agent can discover available capabilities and use the tools your connection exposes. If it needs another account, review the connection proposal. Connecting a service and approving a specific action are separate decisions.
Set up the work you repeat
Create an agent with a specific role and instructions. Select its tools and skills, and add relevant memory for facts that should remain available across tasks.
For a recurring task, configure a routine with its own model, instructions and schedule. It can run without an assigned agent. Assign an agent when you want to use its job description, selected skills or personal memory. Review each result in execution history; connections and action permissions still apply.
Know what happened
Follow the conversation and execution history. Resolve pending approvals when an action requires one. If a run stops, inspect its latest result and the missing connection, permission or credit balance before retrying.
Use a small task to validate a new connection before delegating a larger workflow. An agent is useful when its output is clear enough to review and its permissions match the job.