Over two weeks this fall, I typed “merge it” to my coding agents 42 times. Counting every version of it, 201 of my 1,422 messages were nudges: their only job was getting an agent to push, open a pull request, or merge work it had already finished. Those are the steps that hand finished code back to the shared project.
Every session, meaning every new conversation with an agent, starts by reading a file of standing instructions. One line in mine said “Push after commit when asked.” So every session finished the work, then waited for me to ask. Each “merge it” fixed one session. The next one read the same line and waited again.
That’s my answer to a question I keep hearing: what kind of person is good at AI? My bet is a skill, not a temperament: the written half of management. That means writing down what a newcomer needs, saying what done looks like, and checking for what the writing misses. The other half, coaching people and earning their trust, has nothing to carry it from one session to the next. What’s written reaches every session. What’s said doesn’t.
The practical version is one move, plus a second I’m still testing. When an agent keeps making the same mistake, change what it reads instead of correcting it again in conversation. When it ignores what it reads, add a check that stops it.
Management Is Half the Answer
The obvious answer to “who is good at AI” is managers, and that case has already been made well.
Ethan Mollick’s Management as AI superpower (January 2026) describes an experimental class at the University of Pennsylvania. Executive MBA students, few of whom had ever coded, built working startup prototypes in four days. Mollick’s explanation:
“What’s scarce is knowing what to ask for.”
The students had spent years scoping problems and defining deliverables in their own fields. Those habits became their prompts.
In May, David Curlewis argued that the analogy breaks where managing gets hard. The hard parts of leading people, in Curlewis’s account, are the long ones: careers grown over years, motivation that sags, trust that takes years to earn. None of that applies to a fleet of agents.
Mollick is describing the half of management that can be written down: the scope, the deliverable, the standard for good. Curlewis is describing the half that lives in a relationship.
Only the relationship half needs the other person to grow, and an agent doesn’t grow.
My Coaching Rubric Has No Box for an Agent
Last year I wrote Coach Up or Coach Out about what I watch for in someone’s first weeks on a delivery team. Coach up means invest in them. Coach out means stop carrying them.
The four signals I invest in are curiosity without panic, self-correction, situational awareness, and small wins. Self-correction is the one that matters here: they miss something, and I don’t have to say it twice. One of the coach-out signs is the opposite. They repeat the same misstep because they never reflected on it.
Across sessions, an agent fails self-correction unless the correction is written somewhere it reads. It does fine on small wins. I ship real work with it daily.
That post ends with a choice: coach up the people who are close, and stop covering for the ones who never will be. An agent fits neither side. It doesn’t get closer on its own, because it keeps nothing it’s told. And it’s too useful to coach out.
So the coaching goes somewhere else: into what the agent reads.
The Onboarding Never Ends
In March I wrote that I’d stopped treating the agent like a tool and started treating it like “a team member who needs onboarding,” with the instruction file as the onboarding doc.
That onboarding happens again every session, so the onboarding doc is most of the relationship.
Split by whether a skill can be written down and handed over, the management skills sort cleanly:
| Management skill | Carries over to agents? | Why |
|---|---|---|
| Coaching and growing people | No | It builds a person over time, and the agent starts fresh every session |
| Motivation, trust, psychological safety | No | They live in a relationship, and nothing carries one over |
| Writing the brief | Yes | Same job: give context the reader lacks |
| Defining done | Yes | Done has to be stated, not assumed |
| Checking the work | Yes | A report of done is a claim until someone checks it |
| Writing lessons down | Yes, and it matters most | What’s said in one session is gone in the next |
A Correction in Conversation Lasts One Session
I counted the 201 with a small script that reads my agent transcripts. It picks out the nudges by their shape: short messages like “merge it,” “next,” “keep going.” Then it sorts them by where the agent stopped: before a push, before a merge, before cleanup. A stop that came right after a permission check refused the agent goes in a separate pile, because there the agent had no choice.
On October 9, I replaced the line. The new instruction says the work isn’t finished until it’s committed, pushed, merged, and cleaned up, and to do those steps without being asked.
In the ten days after, nudges to push, open a pull request, or merge fell from about one message in seven to about one in thirty: 8 of 246, against 201 of 1,422.
The same day, I also changed the permission settings, the rules that decide which actions an agent can take without asking me. I let agents merge their own pull requests, and I described opening one as routine to the system that reviews agent actions. So the merge and pull request numbers can’t all be credited to the line.
Pushes are the cleaner case. A standing permission already let agents push without asking before the change, so permissions can’t explain a drop there. Nudges asking for a push went from about 30 in two weeks to none, apart from two that were really new instructions. At the old share of my messages, the second window would still have had about five.
The other limits stay next to the numbers:
- I typed far less overall in the second window, about 25 messages a day against about 100
- The script sorts messages by pattern, so some land in the wrong pile
- It’s one person’s sessions, before and after, not a study
So the push numbers are the ones I’d lean on. About thirty corrections typed in conversation never changed what the next session did. One changed line did.
Writing It Down Isn’t the Whole Skill
The written half has its own failure.
An earlier test, in August, showed it. I added a short brief that loads at the start of every session, alongside the standing instructions, to orient the agent. Ten days later, a hand count of the corrections in my messages had barely moved: 17 corrections in 272 messages, against 14 in 194 before. That’s too few to show an effect either way.
The more useful number was in what those corrections were about. Of the 17, 9 concerned a rule already written into the standing instructions:
- Verify before you claim something works
- Don’t over-engineer
The agent had the rule in front of it and did the other thing anyway. Reaching an agent isn’t the same as being followed. My agents followed “Push after commit when asked” to the letter and the rules I cared about unevenly, and the data doesn’t say why.
For a rule that keeps getting ignored, I’ve started adding a mechanical check: a small program that runs before a risky action and refuses it when the rule isn’t met. The first covers the tool correction I’d repeated most: it stops an agent from publishing a social post unless it has named the publishing route my rules require. I haven’t measured whether the checks work.
So the written half has two parts: the instructions, and a check for each rule the instructions alone don’t hold.
What Would Prove This Wrong
This argument rests on one person’s sessions, so it should make predictions other people can check.
If the written half is the skill, the people who already have it are the ones who’ve had to write for a reader who wasn’t in the room:
- Leads who write handoff notes before they rotate off a project
- Consultants who write the scope before anyone is staffed
- Tech leads who write the runbook for whoever is on call next
They should do well with agents early, whether or not they’ve managed anyone. And strong people-managers who led by being in the room should struggle until they start writing things down.
If great coaches do well with agents without any of the written half, the argument is wrong, and temperament deserves another look.
No Numbers for the Checks Yet
Those are the next counts: whether the checks hold, and whether anyone else’s sessions look like mine.
What I do have is a place to look when an agent keeps doing the wrong thing: what it reads first, and what stops it when it ignores that.
The written line behind those 201 messages was five words long.
