Thinking Is Cheaper Than Doing
Some of what an agent does is reversible and costs nothing to redo. The rest sets, and a mistake there costs a cleanup. Spend everything you can in the cheap half. Review the plan before the agent crosses.
Sort the work by what a mistake costs and the rest follows. Reading, searching, and drafting a plan are free to throw away. Editing and shipping cost real money. Neither undoes itself. The agents I trust know that difference and act on it.
Planning model
Thinking is writing on a whiteboard. Get it wrong, wipe it, and start again. Doing is pressing your hand into wet cement, and a mistake there sets.
So a good agent works the whiteboard first, where reading, searching, and planning cost nothing. It steps to the cement once, when it is sure.
Catch the mistake while it is a drawing. The review before the agent acts is a refund.
When I build agents, the order of the work helps more than a smarter model or a better prompt. Finish the thinking, hand over a plan, and let a person read it before anything real moves. Both halves stay. Only the border between them gets a gate.
01Two halves of the work, with very different prices
Watch a few agents run and every task splits the same way. One half is the whiteboard. The agent looks around and sketches a way forward, and nothing on disk changes. Read the wrong file and you read another one. Wipe and try again.
The other half is the cement. The agent edits files, runs commands, and ships the result. Some of it leaves your machine and reaches other people. Get that wrong and you own a cleanup, or a change nobody can take back.
The guide ran this rule earlier, back when the context filled up. The small cleanup goes before the big one. Planning applies it to the whole task. Spend the cheap move first, and cross to the expensive half once.
02A planning mode is a posture, not a feature
A planning mode draws that line inside the agent. The marker roams the whole whiteboard and never reaches the cement. The agent reads, searches, and hands you a plan without touching a file.
The permissions runtime from earlier supplies the mechanism. A call that only looks passes. A call that changes anything waits until you read the plan. The agent keeps every capability, and the gate decides when it is used. Approve the plan and the mode ends, and the same calls run for real.
03The plan is the cheapest place to be wrong
This took me a while to price. A wrong plan costs one line rubbed out and a better one written. A wrong change costs undone edits, taken-back commits, a hunt for what else broke, and an apology downstream.
So do the judging while it is cheap. The agent shows you the plan before it acts, and you approve it or send it back. Look for the assumption the plan rests on. You catch the bad guess for the price of a rewrite.
The refund
Count the review as a refund. Every wrong plan it catches is a cleanup you never pay for. The fastest route through hard work stops once, to be sure, before the change lands.
04Diving straight in is the expensive habit
Skipping the plan feels faster, and it fools people. Take the review off the line and the path loops back. The agent starts editing, and halfway in the approach turns out to be wrong. Now it undoes the edits and drafts the plan it needed first. The work runs twice, and the second run pays for the first. That circling is the price of acting before understanding.
The cost
Planning costs time, and the cheap half fails in its own way. You can redraw forever to put off the change. A plan can also be wrong and then followed too closely. A quick test finds the hole in seconds. Draw enough to make the change safe. Stop at the line that changes nothing about what you do next.
Decision
Split every task into the cheap half and the expensive half. Put the review on the line between them. Before a real change lands, you must be able to read the plan behind it and say yes. An agent that jumps from your request straight to editing files spends your cheapest chance to catch a mistake. Every plan starts by knowing where you are in a codebase. That is where the cheap move goes next.