Is Software Engineering Dead? Anthropic Got It Half Wrong

By the end of 2026, the title "software engineer" won't exist anymore. That isn't a hot take from a Twitter pundit. That's Boris Cherny, head of Claude Code at Anthropic, telling CNN on May 10. The replacement, he says, is builder. His argument is clean: AI now writes most of the code at his company, humans set the direction, and the old role label has outlived its meaning.

He's right that the title is dying. He's wrong about why.

What Cherny is actually describing isn't a job shrinking. It's a job inflating to absorb every adjacent discipline that used to live next door to it. Engineering isn't being replaced. It's eating product management, design, QA, and ops. Builder undersells the change by a mile. The engineers winning right now aren't the ones who stopped writing code. They're the ones who stopped writing only code.

Before we tear it apart, let's give Cherny's case the strongest version it deserves.

The April 2026 Challenger, Gray and Christmas report shows 83,387 announced job cuts last month, up 38% from March. Technology led every other industry with 33,361 cuts in April alone, bringing tech's year-to-date layoffs to 85,411. CBS News' analysis of the same data puts AI at 26% of the cited reasons. Six months ago that number was a rounding error.

Inside Anthropic, the number Cherny actually cites is that Claude Code now writes the majority of production code at his company. He runs the team. He isn't repeating a marketing line; he's describing what he watches every day. His prediction is that this pattern generalizes and the title decays with it.

The benchmark data supports the shape of his claim. Claude Opus 4.5 hit 80% on SWE-Bench, the test that asks an AI to solve real GitHub issues end-to-end. The follow-up release, Claude Opus 4.7, pushed accuracy higher and brought sub-agent orchestration into the model itself. When Anthropic shipped 10 ready-to-run agent templates for financial services on May 5 — pitch deck builder, KYC screening, credit memo drafting — FactSet stock dropped 8% the next day. Thomson Reuters dropped 5%. The market read the playbook: model plus harness plus skill library is portable beyond coding.

If you read only those numbers, the conclusion looks defensible. The model is good enough, the layoff signal is real, and Anthropic runs on the workflow it sells. The question isn't whether code-typing is being automated. It is. The question is what the job becomes when the typing stops.

This is where the death thesis runs into the rest of the data.

Microsoft's Global AI Diffusion Report, published May 7, 2026, is the most comprehensive AI usage snapshot of the year. It reports that git pushes increased 78% year-over-year globally in Q1 2026. That isn't AI replacing pushes. That's more code reaching repositories every single day than ever before in human history. If AI were killing the engineering role, the artifact engineers produce, namely commits, would be flat or falling. It's accelerating at 78%.

In the same window, US software developer employment reached 2.2 million in 2025, an 8.5% year-over-year increase. By March 2026, employment was running 4% higher than March 2025. The labor stats don't show a profession dying. They show a profession growing faster than the overall labor force during the precise window where AI coding tools became default.

This is the Jevons paradox playing out in software in real time. Making coal engines more efficient didn't reduce coal demand. Cheaper energy created entirely new markets that consumed more coal overall. Cheap code does the same thing. When a non-engineer can vibe-code a prototype on a Saturday, they don't replace engineering. They generate a backlog of half-built products that need a real engineer to test, secure, scale, and operate. Demand for the senior end of the role goes up, not down.

You can watch this play out on Twitter every week. Naval Ravikant's absolutely not dead thread on traditional software engineering racked up over 7,000 likes in 48 hours because every senior dev on the timeline is living the contradiction firsthand. Their employer is using Claude Code. Their team is hiring. Both are true at once.

The Challenger layoff data, which is the death thesis's strongest leg, has a wrinkle the headlines miss. The same April report shows that year-to-date layoffs are down 50% compared to the first four months of 2025. The 38% month-over-month spike is real, but it's a rebound off a record-low Q1, not a structural collapse. Coding work is being reshuffled. Total employment is rising.

Words matter. Builder sounds humbler than engineer, which is exactly why it's wrong.

A builder takes a spec and turns it into a thing. A builder assembles. A builder follows a plan. The word implies that the design is done before the work starts, that the hard part has already happened and the human is the manual labor that closes the gap. That framing happens to be excellent for Anthropic, because it positions Claude Code as the design step and reduces the human to assembly. It happens to be terrible for describing what's actually happening to engineers.

Calling a software engineer a 'builder' in 2026 is like calling a film director a 'cameraman' in 1995: technically true, strategically wrong.

An engineer designs systems. An engineer picks the abstractions, decides where failures are allowed to leak, and trades off the dimensions nobody else in the room can see: latency against cost, consistency against availability, build time against runtime, the dependency you can replace next year against the one you can't. AI models are getting good at writing the code that lives inside those decisions. They aren't getting good at making the decisions themselves.

Every abstraction is leaky and someone has to plug the leaks. Claude Code is a leaky abstraction over the real systems engineers deliver. The cleanest possible proof of this is Anthropic's own product: earlier this year, Claude Code measurably degraded by 67% over a six-week window, and only the engineers who could route around the failure, switch models mid-task, or tighten their specs kept shipping. A builder doesn't have those tools. A builder receives a finished spec and follows it. The actual senior partner in the work is whoever holds the model accountable when the model is wrong.

That role has a name. It's called engineering.

The builder reframe is what you call yourself when you want to keep the part of the work that's getting cheaper and quietly hand off the part that isn't. It's a rebrand designed by the company selling the tools, for the company selling the tools. Engineers who adopt it are agreeing to a smaller version of their own job description right before the market is about to need a bigger one.

Here's the contrarian read of Cherny's claim. He's watching his own engineers spend less time typing code and concluding the role is shrinking. The opposite is happening: the role is inflating to swallow the disciplines it used to hand off to.

Three years ago, shipping a feature looked like this. A PM wrote a spec. A designer mocked the screens. An engineer wrote the code. QA tested it. Ops deployed it. Five people, five handoffs, five months. Today, on a team that uses computer-use agents like GPT-5.4 for repetitive work, the same feature ships like this: an engineer writes a spec the model can execute against, prompts a design model for the UI, runs Claude Code against the spec, generates evals while the AI writes the implementation, and ships through an agentic deploy pipeline that an LLM monitors for regression. Same five jobs. One human seat doing all of them.

That isn't engineering shrunk. That's engineering eating the entire stack.

The skills the role now demands would have been four separate job descriptions in 2023. Spec writing as a programming primitive: the new high-leverage skill is producing requirements precise enough for an AI to execute without ambiguity. Product management used to own this. Engineers now do. Eval design: testing has shifted from write asserts on functions to design adversarial cases that prove a non-deterministic model is doing what you claim. QA used to own this. Engineers now do. Skill and prompt orchestration: composing AI capabilities into a workflow that produces consistent output across runs. UX research and growth used to share this. Engineers now do. Agentic operations: deploying, observing, and intervening in long-running model processes. SRE used to own this. Engineers now do.

The Stanford AI Index 2026 makes the same point with global enterprise data: the companies winning at AI aren't the ones automating their engineering teams away. They're the ones expanding what those teams produce. One engineer orchestrating four agents simultaneously is the new unit of output, and that engineer needs more skills than the 2023 version of the role, not fewer.

That isn't decay. That's promotion.

If you're a working engineer reading this, the practical question is what you change between now and the end of 2026. Specifics.

Get fluent in Claude Code and one alternative. Pick Cursor, Codex CLI, or Cline as your second. The skill isn't use Claude. The skill is knowing when to switch when a model degrades, which it will.

Learn MCP and skill systems. Anthropic's Model Context Protocol is the lingua franca for wiring agents into existing tooling. Writing your first MCP server is now the same career inflection point that writing your first REST API was a decade ago.

Build an evals practice. Most teams still test code paths. The teams pulling ahead test model outputs: they design adversarial cases, score with rubric prompts, and gate releases on eval pass rates. If you can't describe what good output looks like in a prompt, you cannot ship AI-backed software.

Treat spec writing as code. The spec is the artifact the model executes against. Vague specs produce vague output. Tighten yours until a model could run them with zero clarifying questions.

Read what Anthropic ships next. The 10 finance agents announced May 5 telegraph the next coding agents. The pattern of domain-specific skill library plus agent template plus harness is what you'll use to build internal tooling at your own company within six months.

Stop calling yourself a builder. The title still matters. Engineering will still be the title that gets paid.

Cherny isn't wrong about the title decay. Every engineer has had the moment of looking at a finished PR and realizing they typed five percent of it. He's wrong about the trajectory. Builder describes a role that's getting smaller. The actual role is getting bigger. The product manager job, the designer job, the QA job, the ops job, none of those disappear when the AI gets good. They land in one seat, and that seat needs an engineer in it, because every other role lacks the systems thinking required to hold an autonomous model accountable.

By December 2026, the title won't be gone. It'll be the only title left.