Planning strategies: ReAct, Plan-and-Execute, and beyond

Not all agent tasks need the same shape of reasoning. Match the planning strategy to the task; don't force every problem into ReAct or plan-and-execute.

The three canonical shapes

1. ReAct (Reason + Act). Model picks one action, observes, reasons, picks the next. Tight loop. Best for exploration-heavy tasks.
2. Plan-and-Execute. Model produces a full plan up front, then executes it step by step. Best for tasks where the plan is mostly known in advance.
3. Tree-of-Thought. Model considers multiple candidate next steps, evaluates, picks the best. Best for high-stakes decisions where exploration has value.

ReAct in practice

Thought: The user wants the Q3 revenue. I should look that up.
Action: get_revenue(quarter="Q3")
Observation: {revenue: 4200000}
Thought: I have the number. I can answer now.
Action: answer("Q3 revenue was $4.2M")

Strengths: simple, transparent, easy to debug. Weaknesses: no long-term coherence; the model can drift on complex tasks. Can loop if the tool responses don't push it forward.

Plan-and-Execute in practice

Plan:
1. Fetch Q3 revenue.
2. Fetch Q2 revenue.
3. Compute delta.
4. Format response.

Executor steps through. After each step, the plan can be adjusted if the world surprised us.

Strengths: coherent for long tasks; the plan is auditable; cheap execution (plan generated once). Weaknesses: rigid if the world is uncertain; bad plans silently hurt for the whole run.

Tree-of-Thought in practice

For a strategic decision:

Candidate A: Send escalation email.
Candidate B: Request more info from user.
Candidate C: Mark as resolved with caveats.

Evaluate: B has lowest user friction, A has highest resolution likelihood.
Pick: B, but prepare fallback to A if no response in 24h.

Strengths: better decisions on ambiguous choices; more robust. Weaknesses: expensive (multiple generations per decision); overkill for simple tasks.

How to pick a strategy

Hybrid patterns (what real systems do)

Most production agents use plan-and-execute at the task level, ReAct within each step. The plan says "step 3: research competitor pricing." The ReAct loop within step 3 figures out how to do that — searches, parses results, decides when it has enough.

This balances long-horizon coherence with short-horizon flexibility.

What kills planning quality

• Vague goals. Plans are only as good as the goal they're derived from.
• Missing capabilities. The model plans to use a tool that doesn't exist, gets stuck.
• Over-planning. 20-step plans are fragile. Cap your planner at 5-7 steps; let each step be refined during execution.