Plan first, or feel your way?

What this lesson shows

Plan-and-execute is cheap but brittle; step-by-step adapts but costs more.

The question it opens with

Lay out the whole plan up front, or decide the next step each turn? Both can go wrong.

The walkthrough, in the lesson's own words

• Reality hits mid-task. Predict which agent absorbs it.
• Drag the surprise dial. Watch the cheaper line flip at the crossover.
• Pick by predictability, not by which feels smarter.
• Mid-task, the venue cancels: the address the planner baked into step 3 is now wrong.
• Which agent absorbs the surprise without a restart?
• The adapter re-checks before each step, catches the cancellation, and reroutes. The planner walks into the wall first, then recovers by writing a new plan and redoing the work built on the old venue. Both finish; the planner pays for the re-plan plus the redo.
• One task, plan a launch event in five steps, handed to two agents. The planner writes the whole script first, then executes it mechanically. The adapter is the ReAct loop from 6.11 (reason, act, observe): it looks at the result of each step before choosing the next. Same destination, opposite strategies.
• Surprise dial, how unpredictable is the task?
• Low surprise → planning wins: cheap, predictable, no wasted thinking.
• High surprise → adapting wins: it absorbs the change instead of paying to re-plan.
• Real agents do both: they write a plan for you to review, then adapt while executing it.
• The surprise doesn't kill the plan, it invalidates it. The planner can write a new plan, but that costs a second planning pass, and the steps built on the old venue have to be redone. Re-plan plus redo is exactly what the chart charges.
• A known, mechanical chore favors a plan; a messy, novel task favors adapting. This is exactly what plan mode in coding agents like Claude Code gives you, a plan to approve before anything runs, so you spend the cheap, predictable strategy where the task is predictable, and stay in the loop where it isn't.
• An agent lays out a perfect 10-step plan for your launch event, then the venue cancels mid-task and it keeps addressing invitations to the old room. When is planning first the wrong call?
• A rigid plan is cheap on a predictable, well-scoped chore, but a surprise forces a re-plan, and the work built on stale assumptions gets redone. When the task is messy or unpredictable, prefer the adapt loop that re-checks before each step, even though it pays a thinking tax every move.
• Venue gone: the scripted step is stale. Stop, write a new plan.
• Cost versus surprise: the plan line and the ReAct line cross over.
• A toy cost model: invented step costs that show the trade-off, not measurements.

Key takeaway

You can pick plan-first vs. adapt-as-you-go by how much the task is likely to surprise you.

What you can do after this lesson

You can choose between planning up front and adapting step-by-step, and explain the trade-off.

Where it leads: Either way, you choose how much leash to give it.

This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.