Configure an agent around a clear responsibility. Its instructions describe how it should work, while capabilities, plugins, and skills determine what it can use to carry out that role.
Write the instructions
Include the purpose, expected inputs, output format, and boundaries. For example:
“Review supplier proposals against the provided procurement policy. Return a comparison table and unresolved questions. Cite the relevant policy requirement. If a required fact is missing, ask for it rather than filling it in.”
Keep long source documents in Sources or project context so the agent can refer to them explicitly.
Choose capabilities and plugins
Open the agent's configuration and review its available tool and plugin controls. Select what the role needs. A capability listed in instructions is not automatically granted by writing its name there.
For external services, confirm that the relevant plugin is available and the required account is connected. If the agent will run in a project, confirm it is attached to that project.
Add skills and test
Attach skills that describe repeatable procedures. Test both an ordinary request and one with missing or conflicting information. Check whether the agent uses the intended source, produces the required format, and asks useful follow-up questions.
After changing configuration, test in a new conversation rather than assuming an existing conversation has discarded all of its earlier context.
See Plugins and project context.