Programmers No Longer Write Code: When AI Agents Take Over Software Development

From Writing Code to Being a "Dispatcher"

One programmer described a dramatic shift in how he works: his old setup was simple — one screen, one mouse, one keyboard. Now, that same desk has a screen split into ten windows, each one with an AI agent "sitting inside" waiting to be assigned a task. He said he spent an entire day without writing a single line of code — not out of laziness, but because he simply didn't need to.

This isn't an isolated case. Boris Cherney, the creator of Claude Code, revealed that he keeps five terminal tabs open locally while simultaneously running five to ten Claude instances in the browser. Peter Mattis, co-founder of Cockroach Labs, said managing five to ten agents at once feels like the right cognitive load. Dima, an engineer at Linear, put it bluntly: before, one mouse, one keyboard, one screen meant you were physically incapable of doing several things at once — now that constraint is gone.

As a result, the programmer's role is shifting from "writer" to "dispatcher" — issuing instructions to one agent, switching over to check another's output, then switching back.

The Editor's Exit and the Data Behind It

If programmers are no longer typing code by hand, where did the editors go? As various observers have noted, Codex, Cursor, and JetBrains are all evolving toward agentic workflows, and traditional editors are quietly stepping aside.

A compelling data point comes from GitHub: the number of pull requests (PRs) created by agents has already surpassed those created by humans. In other words, AI-submitted PRs on GitHub now outnumber human-submitted ones.

Former Amazon engineer Steve Yegge has categorized programmers' AI usage into eight levels: Level 1 uses no AI at all; Level 4 skips reading code entirely and communicates with agents in natural language; Level 5 discards the editor completely; Level 8 builds custom orchestrators to direct thirty or even hundreds of agents simultaneously. He admitted, with genuine sympathy, that some highly talented engineers are still at Level 1 or 2, carefully reviewing every single line of AI-generated code — "I truly feel for them."

Steve Yegge's eight-level framework deserves a closer look. Levels 1 and 2 represent traditional developers who still write most of their code by hand and treat AI as little more than an advanced search engine. Level 3 starts using AI to generate code snippets, but reviews them line by line. Level 4 stops reading code directly, instead validating functionality through natural language conversations with agents. Level 5 abandons the IDE/editor entirely, working purely in a terminal or chat interface. Levels 6 and 7 begin automating testing and deployment workflows. Level 8 builds a custom "orchestrator" — a program or script that parallelizes the work of dozens or hundreds of AI agents, with the engineer monitoring only overall progress and final results. The significance of this framework lies in what it reveals: AI tool adoption isn't a binary on/off switch, but a continuous spectrum. The productivity gap between levels can be tenfold or more, which is widening the capability divide among engineers.

Code Reviews Nobody Can Keep Up With

Rapidly accelerating output has created a new problem: when ten AIs are writing code at full speed, who actually reviews it?

One analogy captures the situation well: pull requests arrive like packages on a conveyor belt, one after another. You write three thoughtful comments on the first one, give the tenth a quick glance, and rubber-stamp the hundredth without looking. One engineer named what everyone in the industry already knows but won't say out loud — everyone is performing a code review theater. Faced with five to ten times the review volume, nobody can keep up, so people default to the path of least resistance: abandon real review, and approve everything.

The consequences? Tool creator Mario Zechner says "everything is broken" — 98% availability is becoming the norm rather than the exception, even for major services, with strange UI bugs cropping up that should have been caught by QA before shipping.

The irony is that programmers aren't actually working less. Dima notes there's always another agent waiting for your response. At first, he could finish in an hour what used to take a colleague a day — but now everyone assumes you can handle multiple things at once, and whatever time the AI saves gets refilled with more work.

Pull Requests (PRs) are the backbone of modern software collaboration: a developer completes a feature on a separate branch, then submits a merge request to the main codebase. Teammates conduct a code review to verify logic, check for security issues, and confirm the code meets standards — only then can it be merged and deployed. Code review has long been considered the last human line of defense for software quality — it catches bugs, but also transmits architectural thinking and team norms. When AI increases PR submission speed by five to ten times, but reviewers' time and attention can't scale proportionally, that line of defense quietly collapses. Engineers go through the motions of reviewing without truly understanding the code's intent — effectively outsourcing quality control back to the AI itself, which is precisely the output that most needs critical human scrutiny.

Four Things AI Still Hasn't Changed

Despite the upheaval, a closer look at what engineers are reporting reveals four things AI hasn't disrupted — at least not yet.

Team structure remains the same. Anthropic is considered one of the most AI-forward companies around, but Katelyn Less there says they still operate with "two-pizza team" sizes, with each team owning a slice of the software through continuous iteration and on-call rotations — much the same as before.

Planning complex projects still requires humans. Mapping out complex projects still demands rounds of discussion and sometimes digging up documents from two years ago.

Testing rigor hasn't diminished. Jared Sumner, creator of Bun, says AI wrote essentially all the code and all the tests — but the time spent on testing is still roughly equal to the time spent writing code. You still need a way to trust that what was built actually works.

Engineers still make the final call. Product managers and designers still don't ship code directly to production. Nobody surveyed found a single company where that had changed. Whether something goes live is still, ultimately, an engineer's decision.

When You're Exhausted, Remember Why You Started

Faced with this relentless pace, Peter Mattis offers a grounding piece of advice: when you're tired, remind yourself why you got into this field. He got into it because he loved building things — and now, he gets to build faster than ever.

That may be the most understated yet resonant takeaway from all of this. Tools are changing, roles are changing, productivity is changing — but the passion for building things remains the underlying force that keeps programmers going. As for "which level are you at" — that's a question every developer will have to answer for themselves.