Expertise Shifted Left. Where Do the Next Experts Come From?

Published August 13, 2026

A Field Guide entry in “AI-Native Ways of Working.”

Everything I shipped this year was built by agents, and watching it happen taught me nothing about how to do it myself. A year of that, and I can write about as much software as I could at the start, which is none. I'm not an engineer and I'm not becoming one.

That's the trade I meant to make. My job is to aim the work and decide whether what comes back is any good, and the discernment I use for that came from eighteen years of coaching teams, not from writing software. The arrangement works because I brought the discernment in with me.

The cost shows up in the sentence I use most: just do whatever you think is next in the backlog. Months of that on my orchestration layer and on a knowledge base I've been building, and I can see things changing without being able to tell whether the work is making progress. I can read what changed and have an agent walk me through it. What I can't say is whether the shape it's taking is good, because I never did the work that would have taught me what good looks like here.

Debugging is the loss I feel most. When something breaks I hand it back instead of sitting with it, and sitting with it is what teaches you how a system works. My blast radius is also small enough that I've never had to learn production-grade governance. That's a convenience now and a ceiling later.

The question I left open

I ended the last essay by drawing a boundary around this one. What I said there is what I believe: expertise shifts left. Experts put their discernment into designing and validating the system, so the work needs less of it in the moment. Less, not none — I still stop agents mid-run most days. But most of what my agents get right, they get right because of a decision I made before they started.

That belief assumes the experts, and says nothing about where the next ones come from. If the seniors design the system and the agents do the work, the middle of that arrangement — the part where somebody used to turn into a senior — isn't in the picture. When your seniors retire, where does the discernment come from, if nobody has been growing into the role?

Where mine came from

Talking with people about their problems, mostly. Different companies, domains and kinds of problem, far more of other people's than my own. Do that across enough situations and you start recognizing the shape of a thing before the person describing it has finished. I did formal training along the way and it helped, but it isn't where the discernment came from.

One rep I've written about before: getting a team to stop starting things and start finishing them. I could hand you the steps on a page and you'd have the steps, which are not the skill. The skill is reading whether a room's pushback is a real objection or the discomfort talking. Nobody was tracking that as development. The reps came attached to the job, and the discernment was a byproduct. For the person coming up behind me, the work that would carry those reps — the building, the shipping, the finishing — is what agents do now.

The ladder from the other side

In the first of these essays I called earned autonomy one ladder seen from two sides: every rung an agent earns is a rung of altitude I gain. That's also a description of a career, except that a person has to spend real time on each rung. I used to think the bottom rungs were where discernment started. I think now it's the climbing itself.

What I do now is hand the bottom rungs to agents. For me there was nothing down there to learn; I was never going to become an engineer by writing my own migrations. For someone twenty years behind me, that's where most of the learning would have been.

One of the four values I've written for this framework is human judgment where it matters over human hands on everything. I'd write it again, and it carries a debt I didn't name: hands on everything is how most of that judgment got built.

What I see, and what other people are counting

AI-written text has tells: wordy, over-explained, the same rhythms every time. I'm not bothered that someone used AI, since I use it constantly. What bothers me is that nobody read it afterward with a standard in mind. I see it in products too: people who've built an MMORPG or an eSIM reseller and call it a SaaS business, with no grasp of what it costs at a thousand users or what's legal to resell across borders. And I see it in engineers — ten things in flight and no clear answer to how finished any of them are, which is the same gap my delegated backlog opened in me.

Universities have the clearest version, because they got hit first and they grade. Jason Gibson, a history professor at Alcorn State, hid white-on-white text in a midterm PDF telling any AI that read it to work the word “Madagascar” into the answer in a way that made no sense. Thirty-two of thirty-five students turned in a Madagascar.

The starkest number is from Brown, where the economist Roberto Serrano gave a take-home midterm and watched the class average land at 96%, against a historical range of 65 to 80. He moved the final in-person and closed-book, and the average came in around 48%. Same students, same semester, same material. The cheating is the least interesting part of it. That gap is the difference between using AI to get the answer and using AI to learn, and on the day the work is handed in, the two look identical.

The labor version is further along than most companies will say out loud. Stanford's Digital Economy Lab, using payroll records, found a 16% relative decline in employment for 22-to-25-year-olds in the most AI-exposed occupations, software engineering among them, while more experienced people in the same occupations held steady. In LeadDev's survey of 883 engineering leaders, 38% said AI has already cut how much mentoring juniors get from seniors. Matt Garman, who runs AWS, called replacing junior engineers with AI “one of the dumbest things I've ever heard” — they're the cheapest people you employ, and a company with no pipeline eventually collapses in on itself.

The evidence doesn't all run my way. The largest randomized trials, nearly 4,900 developers across Microsoft, Accenture and a Fortune 100 manufacturer, found the less experienced gained more throughput from Copilot. Throughput is what those trials measured. Nobody checked whether the juniors could tell good work from bad, and that is the thing I'm arguing stops getting built.

Learning has to be put back on purpose

The third shift I've written about says AI does the building now, so the learning you used to get while building has to be put back on purpose. I meant that about products.

Applied to people, the same idea costs far more. Learning that used to come free with the work now has to be designed and paid for. I don't know many organizations with that on a plan. My own version of it is the walkthrough from the third essay, and it only goes so far: it hands me back an understanding of what a system does and where it stands, and none of the ability that produced it.

The model I keep coming back to is memory. Agents keep nothing between sessions, so any growth I want from them has to live outside them, in role files and project memory. It's tedious and I'm not good at it yet, and what it forced on me is total explicitness. I've spent a career building growth in people on purpose — teaching, mentoring, coaching, running workshops — but most of that I could do in conversation, trusting the person to keep going after I left the room. An agent gives me none of that. I have to write all of it down. I've started to think the AI era is going to ask for that version with people. When growth stops happening on its own, someone has to build it deliberately or it doesn't happen.

None of us has been at this long enough to have the answer. I'm one person running a lot of agents, with no juniors on the other side of it. This is what I'm seeing. Are you seeing it too?

The anti-patterns

• The atrophied reviewer. The senior person whose whole job is now approving what agents produced. Judging work and doing work are not the same activity, and the discernment that makes the first worth having came from the second. Review keeps that discernment fed for a while, and nobody knows how long.
• Prompt-only juniors. Someone whose entire experience of the craft is describing what they want and accepting what comes back. They'll ship plenty. What they can't do is tell you whether it's any good.
• Hiring your way out. Treating a formation problem as a recruiting problem, on the assumption senior people can be bought when needed. That works until enough organizations try it at once.
• Fluency monoculture. One or two people get good at running agents and everything routes through them. It's one person's expertise doing a team's worth of work, and it leaves when they do.

For leaders

I've said publicly that using AI well takes more people, not fewer, and I still think so. The work doesn't disappear, it moves, and somebody still has to aim it and own the result. That takes discernment, and discernment doesn't come free. This is the uncomfortable half of the same argument: demand for it is climbing, and the process that used to supply it is the one we just automated. It's the same governance question as the rest of this series, and what it asks you to operate is a pipeline that used to run itself.

Two questions get at where you stand. Pick someone a few levels down and ask what they can judge now that they couldn't a year ago, and what they did that taught them; “I got faster” is not an answer. Then ask who owns that answer. In most places nobody does, and nobody had to.

I'm partway on the rest of it, so here's where I'd put my money. Keep hiring juniors; you're buying the seniors you'll need in eight years and there's no other supplier. Aim their growth at discernment and systems thinking rather than at output, because output is the part that just got cheap. And teach working with AI as a skill of its own: what to delegate, how to specify it, and how to judge what comes back.

Four practices I'd run as experiments:

• Deliberate practice. Ericsson's version — reps chosen for what they build in the person rather than what they produce, with real feedback attached. It's the only mechanism I know that manufactures reps a job no longer supplies.
• Verification as a path to mastery. Do the work, then verify it against something real, and let the verifying be where the learning lands.
• Redesigned apprenticeship. The old version handed the novice easy work and let them climb. The new one puts them on review and validation early, next to someone who can say why an answer is wrong.
• Cross-level help-seeking, in both directions. Juniors bringing real problems to seniors to watch how one gets worked, and seniors pulling juniors into problems they're solving right now. The second direction is the one that goes missing.

Shifting expertise left is still the right answer for the experts an organization already has. It says nothing about the ones it hasn't grown. I built my own discernment on work that needed doing anyway, with the learning attached to it, and that deal isn't on the table for the person twenty years behind me. That doesn't mean discernment can't be built now — I think it can, in shapes I never used: simulation, front-loaded discovery, arguing with an agent until you understand the problem. What doesn't survive is expertise arriving by accident. Somebody has to decide it's going to happen and pay for it.

This is one pattern from a set I'm working through on human-agent teams. It comes out of my own hands-on R&D, not a production deployment.

Back to Field Notes