Stop Reviewing Lint (AI Agents & Theory of Constraints)

A recipe from a factory floor

Theory of Constraints says every system has exactly one bottleneck. Improve anything *other* than that bottleneck and you're wasting time. It gives you five steps. In order. It's a simple recipe, but hard to follow because people always want to skip to step 4. This series walks all five, applied to AI coding agents. Let's start at the top.

Step 1: Find the constraint

This dot is you. You review code that AI agents write. Nothing ships until you say so. That makes you the constraint, because the system moves at the speed you can read. In every AI-assisted dev team I've seen, the human reviewer is the bottleneck. Your agents produce more code in an hour than you can review in a day. You will always have more input than you can handle. Step 1: done.

Step 2: Exploit the constraint

Exploit sounds aggressive. It just means: make sure the bottleneck never wastes a cycle. You must be busy doing high-value work. Don't fiddle with tooling. Don't waste your time on nitpicky things. Automate as much as you can. What's eating your review time right now? Let's find out.

Your agents just dropped off 10 PRs

They're fast! Good for them.

Uh oh

9 out of 10 have problems. But wait...

Most of those problems are dumb

Wrong indentation, missing semicolons, someone forgot to run the formatter.

Only 2 have actual bugs — the kind that need your brain.

7 trivial · 2 real bugs · 1 clean

Without CI: all 10 land on your desk

Go ahead, send some through:

So here's what just happened

And 7 of those round-trips? For stuff a computer could catch in 2 seconds.

You — the scarce resource, the only person who can hit merge — spent your afternoon being a spellchecker.

What if we put some cheap robots in front?

A linter and a formatter. Nothing fancy, and these have existed forever.

They check every PR before it reaches you, so the dumb stuff gets bounced instantly. You never see it.

Same 10 PRs. This time with robots.

Try it:

Look at that

19 → 5

You're the same person with the same hours and the same brain. The only difference is less garbage in front of it.

That's Step 2

The linter didn't make you faster. It made sure that when you do review something, it's worth reviewing.

This is boring on purpose. Linters have existed for decades. If your agents submit PRs that fail lint, go set up CI. We'll wait.

The boring stuff matters because it's cheap. You don't need a workflow rewrite or new tools, just a CI pipeline that catches problems a computer solves in 2 seconds.

What everyone tries next

The pitch: "Run 5 agents in parallel! 10x throughput!"

Your agents now produce 50 PRs a day.

You still review 5.

The other 45 sit there, going stale and accumulating merge conflicts.

That didn't increase throughput. It increased inventory — work piling up in front of the constraint, rotting.

Step 3 of Theory of Constraints has a name for the rule you just broke: Subordinate. Don't let non-constraints produce faster than the constraint can consume.

That's a future post.

This series applies the 5 Focusing Steps to AI coding agents. We're in Step 2 — Exploit. Linters are just the start.