Design an AI agent by specifying the decisions it must make and the boundaries within which it may act. Start with a task contract: what arrives as input, what result is required, which changes are allowed, and who accepts the output. Choose an agent only if deciding the next action genuinely benefits from flexibility.
Take a request to check whether a project report contains the required sections. If the section names and checks are fixed, a deterministic checklist may be enough. If the report uses varied wording and the system must inspect supporting documents to resolve omissions, an agent may be useful. This decision comes before selecting a framework or adding memory.
Map the information available at each decision. Specify where the agent obtains it, how current it must be, and what happens when it is missing or contradictory. Then define the permitted actions and the state that must survive interruptions. A report reviewer might need to remember which sections it already checked, without retaining unrelated documents after the task ends.
Finish the design with acceptance and failure cases. A valid result could be a checklist whose findings each cite a report passage; an incomplete result should identify the missing input. Decide when the system should stop, ask a person, or retry. Review these cases with the person responsible for the work. A diagram is useful only if it makes those operational choices visible.
Source: Anthropic, Building effective agents.
