The Green Checkmark Is the Most Dangerous: The Agent Wrote the Tests and They Passed—So Why You Still Shouldn't Rush to Merge

Hesam shared Addy Osmani's article, saying anyone who works with coding agents every day should read it. His original tweet had only one line of accompanying text: there's a fine line between outsourcing work to AI and handing over your judgment.

Addy's article doesn't stop at empty platitudes like 'you need to review code.' The two images list eight items, all actions you can take right away.

The first image has four items: proactively set up the agent to quiz you, interview you, and show you evidence—provided you set it up that way first; be honest about where you are on the spectrum—if you can't explain the answer even with the agent's help, then you've handed it over; form a hypothesis before asking, even if the hypothesis is bad; a green checkmark only proves one thing—ask what it checked and what it didn't check.

The second image continues: look at the evidence first—read what it ran and what it returned—then look at the summary; read which decisions it made for you; how you write the prompt shows how well you understand the problem; set a line and hold it—if you're not willing to sign your name to it, don't ship it.

The comments fill in details. Someone asked where the line is. Someone said, if you turn off the agent and can't explain this diff to someone else, then you didn't review it—you just accepted it. Someone admitted the green checkmark is the biggest trap: the agent writes both the feature and the tests, so a wrong assumption can get written into both sides at once; the tests pass, and nothing has been proven.

John Kovacs put it more concretely: broken PRs usually have someone who can catch them. What's truly dangerous is a PR that looks clean; weeks later, when other code touches it, no one can explain why it was written that way back then. Gao Yang mentioned that the agent rewrote a function he hadn't read, and he only found out in the end because of the old variable names printed on paper on his desk. Someone else joked that some days they're reviewing the agent, and some days they're just its emotional support human.

The comments aren't all one voice. Some netizens think that once models get this strong, the gap will close quickly, and simply following the agent's options can also get you to 95%; such principles will eventually become outdated. That judgment carries its own risk: if you only follow the agent's options, weeks later no one can explain that code, which lands right back on the clean-PR problem John Kovacs described.

The stronger the agent and the smoother the output, the easier it is for people to skip the step they most need to do: a second check. Ordinary users are the same: ask AI to summarize a long article, read only the summary, and you've already handed over your judgment. The only difference is that the diff has become a summary.

Here's one action to take: next time the agent says the tests pass, don't click merge yet. Ask what it checked and what it didn't. Can you turn off the agent and explain this change clearly? If you can't, the green checkmark is just a green checkmark.