The bottleneck moved to review

A plan used to cost an afternoon. You sat with a blank page, argued with yourself, drew a diagram, threw it out, and drew it again. By the time you had something to show a colleague, you had already done a fair amount of the thinking that mattered. The effort of producing the document doubled as a filter. Only ideas you cared about survived long enough to be written down.

That filter is gone. An agent will produce a plan, a spec, a migration outline or a diff while you refill your coffee. The document arrives looking finished. It has headings and a confident tone and a list of risks. Nothing about its polish tells you whether it is right.

This is an opinion piece, and I want to say so plainly: I think the scarce step in agentic engineering is no longer producing the artifact. It is a person reading it, disagreeing with it precisely, and getting that disagreement back to whatever produced it while the context is still warm. Everything else got cheaper. That step did not.

Cheap generation moves the constraint

Here is the reasoning, and you can check it against your own week. When something becomes cheap to make, you make more of it. If plans cost an afternoon, you wrote a few. If they cost a minute, you ask for several variants, then a revision of the best one, then a version that accounts for a constraint you forgot to mention the first time. The volume of things waiting for a human goes up, while the number of humans and the number of hours they have stays where it was.

So the queue moves. It used to sit in front of the writer. Now it sits in front of the reader. You can check this against your own week: the wait is rarely for the model anymore. It is for someone to have time and headspace to sit down with the output and judge it.

Judgment is also harder to speed up than generation. A model can be pointed at a bigger context window or a faster machine. A reviewer cannot be. A reviewer has to hold the goal in their head, hold the document in their head, and notice the gap between them. That is real work, and it does not shrink because the document was easy to produce.

What a review step needs

If review is now where the effort sits, it deserves the same design attention we give the prompt. I do not mean process for its own sake. I mean a few properties that I think make the difference between review that works and review that just happens.

First, the reader has to be able to open the thing without ceremony. Every step between "the plan exists" and "someone is reading it" is a place where the plan sits unread. If the reviewer needs a particular editor, a repo checkout or an agent set up on their machine, I think some reviewers will simply not get to it today, and today's answer is often the one the agent needed.

Second, comments have to attach to specific words. "The second section feels off" makes the writer guess. "This line assumes the queue is ordered, and it is not" can be acted on. A comment tied to a line carries its own context, and it lets the person reading it skip the search for what you meant.

Third, the disagreement has to reach whoever holds the context. This is the one I think gets the least attention. A plan is not just the text. It is also the decisions that went into it, the dead ends that were tried and dropped, and the constraints the author knew about and never wrote down. The session that produced the plan has all of that. A fresh session handed the same file has only the file. When a reviewer's comment goes to a cold reader, someone has to reconstruct the missing context, and reconstruction is where subtle errors come from.

Fourth, the reviewer has to see the revision arrive. If you comment and then hear nothing until someone pings you with "updated, take another look," you will reread the whole document from the top, because you cannot tell what moved. That reread is the expensive part. A review step that shows the change in place, against the comment that caused it, lets the reader confirm the fix and move on.

What goes wrong when review is an afterthought

The failure modes are quiet. Nothing crashes. The work just gets slower and a little worse.

Comments end up in chat, detached from the text they were about. Someone pastes a paragraph, someone replies with a reaction to it, and three messages later nobody is sure which version of the paragraph is being discussed. The writer, human or agent, then has to work from a summary of the feedback rather than the feedback.

Each round starts with a fresh reread. Because the reviewer cannot see what changed, they treat every revision as a new document. This is fair, since they have no way to trust that only the discussed parts moved. But it means the cost of review scales with the number of rounds, and rounds are exactly what cheap generation encourages.

Reviewers never see the outcome of what they said. They leave a comment, get busy, and come back to find a new document. Did their point land? They cannot tell without another full read. Over time I think this teaches people that commenting is not worth the effort, which is the opposite of what you want when their judgment is the scarce thing.

None of this is exotic. It is what happens when a step nobody designed becomes the step everything waits on.

One way to build it

We built Gander around this bet: a reviewer needs a browser and a sign-in, agent connect (start a comment with @agent) reaches the agent session that wrote the file, and the same link updates when the agent edits. It is one example of a designed review step, and a shared document plus a disciplined habit can get you part of the way there. The properties above are the point, whatever carries them.

The question to ask instead

Most of the advice around working with agents is about the writer side. Write a better prompt. Give more context. Break the task down. Some of it is good. But I think it points people at the half of the loop that already got cheap.

The more useful question, if you accept the argument above, is different. Not "how do I prompt this better," but "how does my reviewer's disagreement reach the thing that wrote the plan, and how does the reviewer find out it arrived?"

Ask it about your own setup. Where do comments on an agent's output go today? Who can open them without help? Does the session that has the context ever see them, or does someone summarize them into a new prompt? When the revision comes back, can the person who objected see their objection resolved, or do they start over?

If the honest answers are messy, that is not a criticism of your team. It is what you would expect from a step that used to be a small part of the work and quietly became the biggest part. The prompt got most of the attention because the prompt used to be the hard part. It is worth noticing that the hard part moved, and giving the new hard part the same care.