Plan-execute

Plan-execute agents generate a plan upfront, optionally get it approved, then execute step by step. Trading adaptiveness for predictability.

The flow

1. Given goal G, generate plan [step_1, step_2, ..., step_n]
2. (Optional) Show plan to user, get approval
3. For each step:
   a. Execute (tool call or reasoning)
   b. Observe result
   c. If result breaks assumptions, replan; otherwise continue
4. Return final result

Why plan first?

• Fewer surprises. You see the full plan before any tool call happens.
• Cheaper. Simple steps don't need LLM reasoning between them, the plan already specified what to do.
• Easier to parallelize. Independent plan steps can run concurrently.
• Easier to audit. The plan is the record of intent.

Where plan-execute excels

• Data migrations. Steps are known upfront, no branching needed.
• Multi-file refactors. Plan the list of files to edit, then edit.
• Scheduled reports. Fetch from these sources → process → assemble → send.
• Long-horizon tasks. Agent should keep context, not reconstruct it at each step.

Where plan-execute fails

• Exploratory work. You don't know the plan upfront; you learn it by running.
• High-branching. If half the plan depends on results from the first half, replanning constantly is worse than ReAct.
• User-driven. If the user keeps redirecting, a rigid plan is friction.

Hybrid: plan → execute with replan triggers

Best of both worlds. Plan first, but every step also asks: "does this result violate an assumption the plan made?" If yes, replan that branch. If no, continue executing.

This is what Claude Code's "plan mode" does. The model produces a plan, user approves, execution proceeds, but if something unexpected happens, the model can re-enter plan mode.

Plan quality matters

A bad plan is worse than no plan. Some markers of a good plan:

• Each step is concrete (names a tool, input, expected output)
• Steps are independent where possible
• Risks and unknowns are flagged
• A verification step exists at the end

When to require user approval of the plan

• When the task touches production systems
• When the task is long enough that you don't want to find out the plan was wrong at step 15
• When the task costs money at each step
• First time running a new type of task, build trust before automating approval

Claude Code implementation note

Claude Code's plan mode surfaces the plan in a structured way and requires explicit approval before execution. For anything autonomous and non-trivial, invoke plan mode first. Routine work can stay in auto mode.