ReAct vs Plan and Execute
Two agent architectures for tool using LLMs
ReAct interleaves thought and action turn by turn; plan and execute drafts the whole plan up front, then executes it. Plan and execute is cheaper on long tasks; ReAct adapts better to surprises.
ReAct
Concept →Reason + Act. The model alternates between short 'thought' outputs and tool calls, observing each result before deciding the next step.
Best for: Short workflows where results guide next actions.
Plan and Execute
Concept →First generate an explicit multi-step plan (often JSON), then execute each step via tools, calling back to the planner if steps fail.
Best for: Long workflows where planning upfront saves calls.
At a glance
| Dimension | ReAct | Plan and Execute |
|---|---|---|
| Decision granularity | Per step | Whole plan up front |
| Planning model calls | One per step | One per plan (or replan) |
| Adapts to surprises | Yes (natively) | Only via replanning |
| Cost on long tasks | High | Lower |
| Debuggability | Linear thought/action trace | Plan + execution logs |
| Best for | Short adaptive workflows | Long planned workflows |
Key differences
- 1ReAct decides one step at a time; plan and execute commits a full plan first
- 2Plan and execute uses fewer planning model calls on long tasks
- 3ReAct adapts to surprising observations; plan and execute may re-plan on failure
- 4Plan and execute separates the planning model from the execution model
- 5ReAct is easier to debug (linear trace); plan and execute needs plan + execution logs
In the interview
- Using ReAct for a 20-step process where the plan barely changes
- Using plan and execute for tasks where each step depends on the previous result
- Overlooking that both patterns need retries and guardrails to survive real tools
How to choose
Adaptive and short → ReAct. Long and plannable → plan and execute.
Common misconceptions
Myth: ReAct is always better because it adapts.
Reality: Every adaptive step is another LLM call. On long tasks the token bill adds up fast and plans get lost.
Myth: Plan and execute doesn't need any thinking at execution time.
Reality: Most implementations still replan on failure, so the planner runs multiple times over a real task.
Memory aid
ReAct: think, act, look. Plan and execute: think everything through first, then just act.
Can you combine them?
Hybrid patterns are common: draft a high-level plan up front, then use ReAct within each step. This gets you cheap structure plus adaptive execution.