Workflow

The bottleneck moved, and nobody updated the org chart

Agents write code in minutes, then a human spends days checking it. The verification bottleneck is the slow step now, and a faster agent only makes it worse.

The Reticle team5 min read

On this page

There's a loop every team building with AI agents now lives inside. An agent researches the problem. An agent writes the code.

Then a person opens the app, clicks around, checks the network tab, reads the logs, and decides whether it actually works. Then the feedback goes back to the agent. That third step is the verification bottleneck, and it is where your week goes.

Three of those four steps got much faster in the last two years. One didn't.

A four-step loop: the agent researches in minutes, writes code in minutes, a person verifies it in hours, and feedback goes back to the agent in minutes.

The verification bottleneck: the human became the slow step

When code took days to write, a few hours of manual verification was rounding error. Now the code arrives in minutes and the verification still takes hours, so the math has flipped: the person checking the work is the bottleneck. Not because they're slow, because they're doing by hand what the rest of the loop does at machine speed.

The numbers say the checking is not optional either. In Stack Overflow's 2025 developer survey, 84% of developers use or plan to use AI tools, yet more of them distrust the accuracy of the output (46%) than trust it (33%). The top frustration, at 66%, is answers that are "almost right, but not quite". Every one of those almost-rights is a manual check somebody has to do.

Bar chart from the Stack Overflow 2025 developer survey: 84% use or plan to use AI tools, 66% name almost-right answers as their top frustration, 46% distrust the accuracy, 33% trust it.

And it's not just slow, it's the worst kind of work. Re-running the same flows, eyeballing the same screens, trying to remember what "correct" looked like last time. It's the task most likely to get rushed, and rushing verification is exactly how a false green slips through.

You can't close a loop at machine speed if one quarter of it runs at human speed.

Faster agents make it worse, not better

The intuitive fix is "make the agent better." But a better agent ships more changes, more often, which means more to verify, not less. A faster agent is a faster way to fill your review queue.

Improving the fast steps widens the gap at the slow one. The bottleneck doesn't move on its own; it just gets more expensive to ignore.

The only way to actually close the loop is to make verification run at the same speed as everything else, automatic, on every change, with a verdict the agent can read and act on without a human in the path. That step is what we mean by AI verification.

What that looks like

The agent makes a change. Verification runs against the real app, real interactions, real network calls, real state, and returns a verdict pointed at the exact line if something broke. The agent fixes it and re-verifies. The human reviews the result, not every intermediate step.

That's not removing people from the loop. It's removing people from the part of the loop that never needed a person, and giving them back the time to do the part that does.

The org chart still says a human signs off on quality. That part's fine. The part that needs updating is the assumption that signing off means doing the checking by hand.

If you want the practical setup, Claude Code and Codex for testing walks through it. If you want the night this problem became a product, that is the founding story.

Fun fact

The best-known lesson on bottlenecks is a scout called Herbie. In Eliyahu Goldratt's business novel The Goal, a troop on a hike keeps stretching out however fast the leader walks, because the slowest boy, Herbie, sets the real pace. They put Herbie at the front, share out the heavy gear in his pack, and get to camp in record time (summary at MindTools). Every review queue has a Herbie. It is usually a person with a browser tab open.

Conclusion

Agents made research and coding fast and left manual verification at human speed, so that one step now sets the pace of the whole loop. A better agent adds changes to the queue. It does not shorten it.

The next step is small. Pick one flow you check by hand after every agent change, and make that check run on its own against the real app, with a verdict the agent can read.

FAQ

What is the verification bottleneck?

The verification bottleneck is the point in an AI agent workflow where a person checks the agent's work by hand. Research and code now take minutes while manual verification still takes hours, so the human checking step sets the pace of the whole loop. Faster agents widen that gap rather than closing it.

Why don't faster AI agents fix the verification bottleneck?

Faster AI agents ship more changes, more often, so there is more to verify, not less. Improving the fast steps only widens the gap at the slow one. The bottleneck closes when verification runs at the same speed as the rest of the loop: automatically, on every change, without a human in the path.

Do developers trust AI-generated code?

Not fully. In Stack Overflow's 2025 developer survey, 84% of developers use or plan to use AI tools, but 46% distrust the accuracy of the output and only 33% trust it. The top frustration, named by 66%, is answers that are almost right, but not quite, which is what manual checking has to catch.

How does manual verification let a false green through?

Manual verification is repetitive: re-running the same flows, eyeballing the same screens, trying to remember what correct looked like last time. That makes it the task most likely to get rushed, and a rushed check is how a false green slips through, with every check marked passed while the change is broken.

Why Reticle

Your agent says it works. Reticle checks.

  • Checks each change inside the running app: network, state and console, not a screenshot.
  • Returns a pass or fail with the evidence and the file:line to fix.
  • Open-source SDK on your machine, or the platform on every preview and PR.

No credit card required.