Planning loops

A ReAct agent makes decisions one step at a time. A planning agent thinks about the whole problem first, produces a plan, then executes. The plan gives the system structure; the execution stays flexible.

The plan-then-execute pattern

1. Plan step: agent is prompted to produce a step-by-step plan for the task. Output is a structured list of actions.
2. Execute step: orchestrator walks the plan, calling tools at each step. Agent can call for replanning if something changes.
3. Final step: agent synthesizes results into a final answer.

When planning helps

• Tasks with many steps where the full shape is knowable up front
• When you want to review the plan before execution (safety, cost control)
• When steps are independent and parallelizable
• When failures mid-task benefit from a "restart from step N" recovery

When planning doesn't help

• Exploratory tasks where each step's result changes the next step radically
• Tasks where you genuinely don't know what the plan looks like until you start
• Low-latency use cases (planning step adds one LLM call)

Hybrid: plan + react

Most modern systems use a hybrid: produce a high-level plan, then run ReAct within each plan step. Plan gives structure, ReAct handles tactics.

Replanning

Real tasks don't go according to plan. Build replanning triggers:

• Tool returned unexpected result
• Cost exceeded budget for a plan step
• User intervened with new info
• Multiple steps failed

On trigger, agent generates a new plan from the current state.

Plan quality is prompt quality

A good plan-step prompt tells the model:

• The high-level goal
• Available tools
• Constraints (budget, time, safety)
• Format for the plan output
• Instructions to prefer fewer steps over more