Ten PRs a Day No Longer Reveal Who Is a Good Engineer
When AI coding agents take over execution, PR counts become noise. RonnyHe, R&D lead for MiniMax Code, breaks judgment criteria into three things: what has been rejected, how much rework after release, and whether constraints can be reused. The replies add a sharp point: if judgment isn't written down, it resets to zero the next day. Salary data at the end.
Two engineers each submit ten PRs (pull requests) on the same day, and it may be two completely different things.
One kind of person is genuinely driving the agent: tasks are broken down cleanly, constraints are written up front, every PR has been reviewed, and after merging no one has to touch it again. The other kind of person sends out whatever the agent wrote directly; the PRs look tidy, but problems surface in review or after launch, and someone else ends up cleaning up the mess.
Looking only at output, the two people are identical. The difference is all below the surface. This is the starting point in the article by RonnyHe, R&D lead for MiniMax Code. His tweet compresses the conclusion even further: when agents write code, ten PRs a day proves almost nothing.
The article is titled Leaders Who Don't Use AI Are Being Left Behind, and its body focuses on how teams identify people and how leaders exercise judgment. He writes that after AI coding became widespread, it has become increasingly hard for him to tell whether someone is a good engineer. The old yardstick was simple: who writes good code, who writes it fast, who can take on the hardest modules. Execution was hard currency. Now that yardstick has stopped working.
Once execution becomes cheap, "can write code" is no longer scarce. Output cannot demonstrate ability; sometimes it even hides problems.
RonnyHe breaks the change in individual value into three layers. The first is balancing quality and speed. Agents make speed easy; the hard part is knowing when to slow down: which changes can be handed to an agent, and which ones you must read yourself. The second is writing constraints for your own agent. Task boundaries, acceptance criteria, and no-go zones must be written down first, or the agent will drift. With good constraints, one agent can do a lot; with bad constraints, the more it does, the more you have to clean up. The third is integrating the output of multiple agents. Do the interfaces line up, do the assumptions conflict, and is the final result a coherent whole? Someone always has to look at the whole.
Behind this is a shift in thinking: from execution thinking to systems thinking. In the past you asked, "How do I write this piece of code?" Now you ask, "How do I split it up, constrain it, check it, and put it back together?"
So how do you tell who is good? RonnyHe offers three questions.
What has he rejected from the agent? Someone who can say "I didn't want this part because..." has at least read it and knows what good looks like. Someone who accepts everything cannot answer.
How much rework is there after release? Shipping fast is not hard; shipping without having to come back and fix it is hard. Rework says more than any output number.
Can his constraints be reused by others? Can the rules, acceptance checklists, and instructions he writes for the agent be picked up and used directly by someone else? If they can be reused, it means he is thinking about a class of problems, not just the task in front of him.
The replies push this standard one step further. One user asked, how do you see rejection? PR counts don't show it. Another reply said, throughput is easy to count; rejection is the signal. John Rood added a harder point: judgment does not survive a session; rejection only compounds if it is written into rules that the next round of agents must read. Otherwise every morning you have to earn the same taste all over again. RonnyHe agreed. He said he looks at review comments, at the constraints someone writes for the agent, and at how often the work comes back after launch; the real signal is what they refuse to merge, and whether that rejection becomes a rule that must be followed in the next round.
This discussion gets closer to the core of the problem than "ten PRs" does. If judgment stays only in someone's head, it is hard for a team to accumulate it. Only when it is written down and turned into constraints can it be reused, checked, and inherited by the next round of agents.
For leaders, RonnyHe's judgment is even less polite. In the past, leaders spent most of their time on short- and medium-term planning, long-term strategy, and building the team, while technical details were left to the team. With enough experience, their judgment would not be too far off. The AI wave has flipped this allocation. Tools and practices are changing so fast that what was learned a few months ago may already be obsolete. With experience expiring this quickly, where does judgment come from? His answer is that you have to have tried it yourself.
If you have not had an agent run a complete task end to end, it is hard to judge whether "this can be automated" is true or false, and it is also hard to know what one person shipping ten PRs a day actually means. The cost of skipping hands-on work shows up quickly: the new tools the team talks about, the pitfalls they hit, and the new ways of working they discover—the leader cannot keep up with any of it. Gradually, the team leaves the leader behind.
He gives leaders three new lines of work: think, build, and go outward. Think means looking at how business and technology are changing in the AI wave, which things people no longer need to do, which things matter more, and whether the shape of the team should change. Build means trying the latest tools and practices yourself, finding support for your judgment, speaking the same language as the team, and writing a bit more code along the way. Go outward means dropping the "I'm just an engineer" mindset and proactively learning GTM and building a personal brand.
The third is the one most easily skipped by leaders with technical backgrounds. Once execution becomes cheap, the thing itself becomes less and less scarce. Team value increasingly depends on being seen: do users know about it, do they understand it, do they want to use it. The same is true for engineers; value needs someone to articulate it. Part of a team's influence depends on whether the leader can clearly explain what the team is doing. If you don't explain it, no matter how good the team is, the outside world won't see it. Learning GTM and building a personal brand are, here, a leader's responsibility to the team; it has little to do with becoming famous.
The market is already pricing this. SignalGigs posted AI training job data from November 11 in the replies: senior developers who use coding agents daily are paid $100 to $200 per hour to judge agent output; across 216 hourly coding jobs, the median hourly rate is $75; and one coding contract uses a 2-hour test to screen people and pays $100 for passing the test.
The scarce skill is saying no to a seemingly reasonable PR, and the premium flows there too.
Agents will keep getting stronger, and execution will keep getting cheaper. Whether you are an individual contributor or a leader, the same question will keep coming up: where does value actually come from? RonnyHe does not give a standard answer at the end of the article; he leaves only one question: how do you now judge whether someone is a good engineer?
发布时间: 2026-11-17 14:27