Optimizing Agents Depends on the Task Design

The article discusses the common misconception that optimizing agents means tuning the underlying model. Instead, the author argues that agent optimization is a task design problem, where the focus should be on defining clear objectives, feedback loops, and constraints rather than just prompt engineering.

Details

The article explains that an agent is not just a language model, but a system composed of various components like task definition, action space, constraints, feedback mechanisms, and stop/escalation conditions. If these components are poorly designed, no amount of prompt tuning will make the system reliable. The author emphasizes that most agent failures are task design failures, where objectives are too broad, success criteria are vague, or responsibilities are overloaded. Effective agent systems rely on feedback loops that are timely, aligned with real objectives, and capable of triggering escalation. Constraints define the boundaries within which the agent can operate, and they provide structure rather than limiting performance. Humans are best positioned to optimize the decision space, not individual decisions, by defining goals, constraints, and interpreting ambiguous situations.