Agent Configuration
Tune prompts, model settings, and attached tools for each agent.
Good configuration makes an agent easier to route, easier to supervise, and harder to hallucinate.
Prompt Shape
Keep prompts operational. State the agent's job, what counts as success, what it must not guess, and how it should behave when evidence is missing.
Runtime Settings
Use model, timeout, and provider choices that fit the task. A low timeout on a heavy reasoning role will look like a logic bug when it is actually a runtime mismatch.
Tool Boundaries
Attach tools only where they are needed. Tool access should flow from the domain, not from convenience. If an agent should summarize but not execute, leave execution tools off the role entirely.
Caution
Validation
After editing an agent, test it with one direct chat prompt and one supervisor-backed task. Then inspect the execution trace to confirm the role behaved the way you intended.
Next Steps
Move to Tool Basics if the role needs new capabilities, or Building Workflows if the role is ready to be reused in a repeatable run.