"Looks good" is the most expensive comment

Here's a hypothetical, made up to show the moment. An agent writes a plan to add rate limiting to a public API. It's tidy: a token bucket per key, a new middleware, a config block, a rollout in three steps, and a section near the end about backfilling usage history from old logs so limits apply from day one. You read it on your phone between meetings. Nothing jumps out. You type "looks good" and move on.

Three steps later, the agent is deep into the backfill. It has written a script that reads months of request logs, and it's asking for credentials to the log store. Nobody wanted the backfill. You skimmed past that section because the first half was right. But you said it looked good, so as far as the agent knew, every part of the plan was approved, including the part you never read.

This is an opinion piece about reviewing agent work, and here's my claim. I think vague approval on agent output costs more than a specific line of disagreement. "Looks good" tells the agent nothing it can act on, and it tells the next person who reads the thread nothing about what was checked. A useful comment names something: a cut, a risk, or a goal you're deliberately not pursuing.

Why "looks good" is expensive

The comment itself is free to write. That's the problem. It costs the writer nothing and hides almost everything.

First, it hides what was read. "Looks good" sounds the same whether you read every line twice or glanced at the headings. The agent can't tell the difference, and neither can a teammate who opens the thread later and sees your approval. They'll reasonably assume someone checked. In the hypothetical, a colleague who saw "looks good" under the plan had every reason to think the backfill was intended.

Second, an agent treats it as license for the whole document. A person who gets "looks good" might still pause on an odd section and ask about it. I think an agent is much more likely to take approval at face value and carry out everything in scope, because that's what approval means. If the plan contains something you didn't want, a general yes turns it from a proposal into an instruction.

Third, the cost lands later and somewhere else. The time you saved by not writing a real comment comes back as rework three steps in, after code exists, when changing direction means throwing something away. The comment was cheap at the moment you wrote it and expensive at the moment somebody had to act on it.

Three comments worth writing instead

You don't need long reviews. You need comments that each carry one specific decision. I think three kinds cover most of what matters on an agent's plan.

The cut. "Drop the backfill section. Limits start fresh at rollout." A cut removes something from scope, and I think it's the comment most likely to save real work, because agents tend to include anything plausibly related to the task. It's also easy to act on. There's no ambiguity about what changes.

The risk. "This touches the billing middleware. Check how the limiter interacts with metered usage before step two." A risk comment points the agent at the place where being wrong would hurt. You may not know the answer yourself. You just know where the danger is, and saying so turns a confident plan into a careful one in the right spot.

The missing non-goal. "Add a line saying we won't change limits for existing enterprise keys." Some of the most important parts of a plan are the things it won't do, and I think agents rarely write those down unprompted. A non-goal stops scope from growing quietly, and it gives the next reviewer something concrete to hold the work to.

Each of these is a sentence or two. None needs a meeting. And each one leaves a record of a decision someone made, which "looks good" never does.

Approval that carries information

None of this means you should never approve. Plenty of plans are fine, and saying so is useful. The fix is to make approval say what you checked.

"Rollout steps and config look right. I didn't review the backfill section." That's still an approval, and it's still short. But now the agent knows which parts are cleared and which aren't, and anyone reading later knows exactly what was looked at. If something goes wrong in the backfill, nobody has to wonder whether it was reviewed. It wasn't, and that's on the record.

"Approved. I checked where the middleware sits against the billing path, and it runs after metering, which is what we want." Here the approval carries the one thing the reviewer cared about. It's a small addition, and I think it changes what the comment is worth. A bare yes is a signal. A yes with a reason is evidence.

We've been building a review loop of our own, and the comments that help most are the ones that say what to change or what was checked, not just whether the reader liked it.

There's a side benefit. Writing down what you checked makes you notice what you didn't. If you go to type "I checked the rollout" and realize you skimmed it, the honest options are to read it properly or to say you didn't. Either is better than a general yes that implies you did.

When you really don't have time

Sometimes the honest situation is that you can't review the plan properly right now. That's fine. Say so. "Haven't read this closely. Don't start the backfill until someone has." is a better comment than "looks good," because it tells the agent where to stop and tells your teammates the review isn't done.

What I'd avoid is the polite middle ground, where a quick skim gets typed up as approval because saying nothing feels rude. Between people, that's a social nicety with a modest cost. With an agent, I think it's an instruction to proceed with everything, including whatever you didn't see.

The cheapest comment to write

Here's the reframe I'd leave you with. We tend to judge review comments by how much effort they take to write, and by that measure "looks good" is the most efficient comment there is.

I think the better measure is what the comment costs the person who has to act on it. Judged that way, "looks good" is expensive. It hands the agent a decision you didn't make, hands your teammates an assurance you didn't earn, and pushes the real review to the point where changes cost the most. A specific line, even "I'd cut this," takes a few seconds more to write and saves the receiver from guessing.

The cheapest comment to write is the most expensive one to receive. Write the one that says something.

Written by the team building Gander, live review for markdown an agent is still writing.