Start with the smallest useful tool set
When creating an agent, pass the specific tools it needs instead of automatically enabling every available tool. A research task may need search and page retrieval, while a local file task may need file operations. Remove capabilities that are unrelated to the requested outcome so the model has fewer actions to choose from.
Tool selection does not replace permission controls. A tool can be available to the agent and still require approval before a sensitive action, depending on the runtime and permission mode. Review both the declared tool set and the approval behavior when designing an agent that can change files, contact services, or perform other consequential work.
Review the scope when the task changes
If you later ask the agent to do work outside its original purpose, check whether it has the appropriate tool and whether adding that tool is justified. Do not make the agent broadly capable merely to avoid a small configuration change; an explicit, narrow tool list makes the agent's scope easier to understand.
For troubleshooting, compare the tools configured for the run with the action the agent attempted. A missing tool can explain why it could not perform a step, while an unexpected tool may indicate that the task or agent definition should be narrowed. Update the tool list deliberately, then review the next run's actions.