Agentic AI & agents
AI Agent Planning
The reasoning layer that decides what an agent should do next, and why.
Planning is the part of an agent that decides what to do next, given a goal and everything it has learned so far. It is what separates an agent from a system that can only call one tool in response to one request — planning is what lets it break a large goal into a sequence of smaller, dependent steps.
Good planning does not mean producing a perfect upfront plan and executing it blindly. In practice, the most reliable agents re-plan continuously, adjusting as each step’s result changes what they know.
Key takeaways
Two planning styles: upfront vs. adaptive
An upfront plan asks the model to lay out the full sequence of steps before taking any action, then execute that sequence. It is easier to inspect and reason about, but breaks down when an early step returns an unexpected result that invalidates the rest of the plan.
An adaptive, step-by-step approach asks the model to decide only the next single action at each point, using the latest results. This handles surprises far more gracefully, though it costs more model calls and can be harder to predict in advance.
Why real agents usually blend both
Most production agents sketch a rough plan upfront — enough to estimate scope and catch obviously impossible goals early — then re-evaluate and adjust after each step rather than following the original plan rigidly.
This hybrid gets the predictability benefits of upfront planning without the brittleness of assuming the first plan will survive contact with real data.
Common planning failure modes
Agents can plan a technically valid sequence that is wildly inefficient, repeating similar steps or exploring dead ends the model should have recognized earlier. They can also fail to notice when a step’s result contradicts an earlier assumption, and continue executing a plan that no longer makes sense.
Logging the full plan and each revision — not just the final action taken — is essential for debugging these failures after the fact.
Put this into practice