06.04 · Concept
When planning makes it worse
Spot the tasks where a plan adds latency and error without adding accuracy.
Planning can make simple agent tasks slower and less reliable when the model could answer directly. For straightforward transformation, classification, extraction, or short-answer work, extra planning creates more calls, more intermediate outputs, and more chances to drift from the user’s request without improving accuracy.
What this lesson answers
- when should an AI agent skip planning
- why can agent planning reduce accuracy
- how to choose direct call versus planning
Notes
Planning is useful when a task has real structure: dependencies, hidden requirements, multiple tools, or a high cost of doing the wrong thing. It becomes harmful when the task is already simple enough for the model to answer directly, because every extra planning step is another model call, another chance to misunderstand, and another delay before producing value.
A good mental model is a production request path. If a user asks for a straightforward transformation, classification, short answer, or extraction, routing it through “make a plan, critique the plan, execute each step” is like…
Common questions
- When is planning useful for an agent task?
- Planning helps when the work has real structure: dependent steps, missing information, tool coordination, or a meaningful cost for choosing the wrong action. In those cases, decomposing the task can reduce uncertainty, make recovery easier, and improve the chance that the agent handles the request correctly.
- Why can a plan make a simple task worse?
- A generated plan is another model output, not a source of truth. For simple work, it can introduce misunderstanding, split the request into unnecessary subproblems, and delay the answer. The agent may then optimise for following its own plan rather than satisfying what the user actually asked.
- How do I decide whether to add planning?
- Ask whether planning will reduce uncertainty or just add ceremony. Use a direct model call for low-risk, single-step tasks such as extraction, classification, simple rewriting, or short answers. Add planning only when it improves accuracy, recovery, tool use, or clarity enough to justify the extra failure points.