Agent planning is how an AI agent breaks a goal into steps, decides their order and revises them as results come in. A request like “find out why our conversion rate dropped last month” has no fixed procedure; the agent must decide what to check first, what that implies, and what to do next. Planning quality largely determines whether an agent completes complex tasks or wanders.
Key takeaways
- Planning can be implicit, decided step by step, or explicit, written out first.
- ReAct interleaves reasoning and action at each step.
- Plan-and-execute writes a plan upfront, then carries it out.
- Reflection lets agents critique and retry their own work.
- Reasoning models improve planning noticeably.
Agent planning explained in simple terms
Given “organise the office move”, one person starts by making a checklist and working through it, adjusting as they go. Another just starts doing the most obvious thing and figures out the rest as they learn. Both can work; both fail differently. The first can follow a bad plan too long; the second can lose track of the goal. Agents face exactly this trade-off.
Main planning strategies
| Strategy | How it works | Suits |
|---|---|---|
| ReAct | Reason, act, observe, repeat, one step at a time | Exploratory tasks with uncertain paths |
| Plan-and-execute | Produce a full plan, then execute steps | Tasks with predictable structure |
| Decomposition | Split the goal into subtasks, solve each | Complex, separable work |
| Hierarchical | High-level plan with sub-plans per step | Long, layered projects |
| Reflection | Critique results and retry weak steps | Quality-sensitive output |
| Tree search | Explore several options and pick the best branch | Problems with clear evaluation |
Many production agents mix these: an initial plan, ReAct-style execution, and reflection at the end.
A worked example
Goal: “Prepare a briefing on our top competitor’s new pricing.”
- Plan: identify the competitor, find their pricing page, compare with ours, check recent announcements, summarise implications.
- Execute step 1: search internal CRM for the named competitor.
- Observe: two products exist, not one. Replan: cover both.
- Execute: fetch both pricing pages, extract tiers.
- Reflect: the comparison lacks our own current prices; retrieve them and redo the table.
- Finish: produce the briefing with sources.
Steps 3 and 5 are where planning earns its value: adapting when reality differs from the plan.
What makes planning fail
| Failure | Cause |
|---|---|
| Over-planning | Long plans made before any information is gathered |
| Plan rigidity | Continuing a plan that new evidence contradicts |
| Goal drift | Losing the original objective over many steps |
| Infinite loops | Repeating the same failing action |
| Premature completion | Declaring success without verifying |
Practical techniques
- Re-state the goal in context periodically.
- Require explicit success criteria before execution starts.
- Cap steps and detect repeated actions.
- Ask the agent to verify results against the criteria before finishing.
- Keep plans short and revise them rather than writing exhaustive ones upfront.
- Use structured output for plans so they can be inspected and tested.
Limitations
- Planning quality falls as tasks lengthen.
- Agents are poor at estimating whether a plan is feasible.
- Verification is often weaker than generation, so errors slip through.
- More planning means more tokens, cost and latency.