Should we always require approval before execution?

For anything that spends, sends or changes something, yes, and for the rest it becomes friction that gets switched off. The useful setting is approval on the steps that are hard to undo rather than on all of them, since a gate people click through without reading is worse than no gate at all.

Is a longer plan a better plan?

No, and the reason is that a plan has two costs rather than one. Every step is work to carry out and work for somebody to read, so past a certain length nobody checks it and the reviewability that justified planning has quietly gone. The right length is what a person will genuinely read.

Glossary

Planning

Planning is a system working out its sequence of steps before carrying any of them out, which produces something a person can read and correct while correcting it is still cheap.

In plain terms

Working out the steps first, then doing them. It matters less because the thinking is better and more because the list exists: somebody can read it and say no before any of it happens. Once the work has started, stopping it costs a great deal more than changing a line did.

Why it matters

Because it is the one moment when correcting course is nearly free. A written sequence can be read in seconds and changed with a sentence, while the same correction after execution means undoing work, spend and whatever the steps already touched. Products that show you the plan are offering that moment; products that do not have decided on your behalf that you did not need it.

How it works

The sequence is produced before anything runs, which is what makes it inspectable. A system asked to do something substantial first states what it intends to do and in what order, and that statement exists as text rather than as an intention buried in a run. Everything useful about the idea follows from it being readable.

Plans meet reality and change, so the first one is rarely the one that executes. A step fails, a page is not where it was expected, a file turns out to be a different shape, and a capable system revises rather than continuing down a route that no longer applies. Watching how a product handles that is more informative than reading its initial plan.

It is the review point products differ most on, and the difference is easy to miss. Some show the plan and wait, some show it and proceed, and some plan invisibly and present only the result, which are three genuinely different products wearing similar descriptions. The question worth asking is whether approval is offered before the expensive part.

Longer plans are not better plans, and depth costs on both sides of the transaction. Each step is work to carry out and work for somebody to read, so a plan nobody has time to check has quietly given back the reviewability that justified having one. The useful length is whatever a person will actually read.

When a correction costs almost nothing

The reason this gets undervalued is that the left-hand column looks like paperwork and the right-hand column looks like the actual work. Reading a numbered list of intended steps feels like an interruption, especially when the system has been right the last several times, and the natural drift is towards skimming it and then towards clicking past it. That drift is rational in the short run and expensive exactly once. What makes it worth resisting is not that the system is often wrong but that the asymmetry is enormous: the cost of reading is a few seconds every time, and the cost of not reading is bounded only by what the system was allowed to reach. Teams that keep the habit tend to be the ones that narrowed the plan rather than the ones with more discipline, because a plan of four steps gets read and a plan of forty does not.

Seen in the wild

An autonomous general agent that plans and executes end to end in its own cloud space, briefed like a contractor and reviewed on what comes back.
Manus
An editor whose agent layer lets a plan be reviewed before a build starts, with parallel sub-agents splitting the work once it does.
Cursor
A cloud development platform whose agent plans multi-step builds and runs tasks in parallel, offering build modes that trade cost against capability.
Replit

Common misconceptions

People assume

The plan is where the intelligence is.

In fact

The plan is where the visibility is, which is more useful to a buyer. Producing a sensible sequence is not the hard part of autonomous work; noticing partway through that the sequence no longer fits is. Judge a product on its revision behaviour rather than on how impressive its opening list looks.

People assume

Seeing a plan means being asked to approve it.

In fact

Those are separate behaviours and products mix them freely. Displaying the intended steps and pausing for a decision are different features, and a system that shows its plan while proceeding immediately has given you a receipt rather than a choice. Establish which one you have before an unattended run.

Telling them apart

Planning vs Task decomposition

Planning

Working out the order, before anything runs.

Task decomposition

Breaking one large job into smaller ones.

Decomposition produces the pieces and planning arranges them, so a system can be good at splitting work and still attempt the parts in an order that wastes most of it.

Questions

Why does a plan matter if the system revises it anyway?: Because the first version is the cheapest thing you will ever be shown. Reading it takes seconds and correcting it takes a sentence, while the same correction after execution means undoing work that has already touched files, systems or a budget. Revision afterwards is the system's job; the plan is yours.
Should we always require approval before execution?: For anything that spends, sends or changes something, yes, and for the rest it becomes friction that gets switched off. The useful setting is approval on the steps that are hard to undo rather than on all of them, since a gate people click through without reading is worse than no gate at all.
Is a longer plan a better plan?: No, and the reason is that a plan has two costs rather than one. Every step is work to carry out and work for somebody to read, so past a certain length nobody checks it and the reviewability that justified planning has quietly gone. The right length is what a person will genuinely read.