Boris Cherny, the Anthropic engineer who built Claude Code as a side project in late 2024, told Y Combinator's Lightcone Podcast this week that coding is "practically solved" for him personally, and that he expects the same will be true for everyone else by the end of 2026. The roughly 50-minute podcast episode, published Tuesday, covers everything from CLAUDE.md files to agent swarms, but the headline-grabbing claim is more existential: the job title "software engineer" is on its way out.
"I think we're going to start to see the title software engineer go away," Cherny said. "And I think it's just going to be maybe builder, maybe product manager, maybe we'll keep the title as a vestigial thing." That's a confident prediction from someone who hasn't manually edited a line of code in months, according to his own account. Whether it's confidence or salesmanship depends on how seriously you take internal Anthropic metrics as representative of the broader industry.
The productivity question
Cherny's personal numbers are hard to argue with, at least on the surface. In a 30-day stretch he shared publicly, he landed 259 pull requests with 497 commits, roughly 40,000 lines added and 38,000 removed, all generated through Claude Code paired with Opus 4.5. He runs five terminal instances simultaneously, each in its own git checkout, with another five to ten sessions in the browser. It's less "developer using a tool" and more "air traffic controller managing a fleet."
Anthropic's internal data tells a related story, though the numbers are messier than the original source material suggests. According to reporting by Pragmatic Engineer, Anthropic saw a 67% increase in PR throughput as engineering headcount doubled, a period that coincided with Claude Code adoption across the org. In a normal scaling scenario, per-engineer output drops as teams grow. That it went up is notable, though attributing the gain entirely to Claude Code requires ignoring every other variable.
The original source for this article claims a "150% productivity increase" and suggests Anthropic engineers are "1,000x more productive than a Google engineer at peak." I couldn't verify either figure in any published interview or reporting. The 67% PR throughput number from Pragmatic Engineer is the most concrete data point available, and it measures something narrower than "productivity."
Build for the model that doesn't exist yet
The more interesting thread in the conversation isn't the bold predictions. It's Cherny's advice to founders: stop building for today's models. His argument is that scaffolding around a model's current limitations can yield 10-20% performance gains, but the next model release often makes that scaffolding irrelevant. You either invest in temporary workarounds or wait six months and get the improvement for free.
"Beginner's mindset is key as the models improve," he says, which is a polite way of saying that the elaborate prompt engineering and CLAUDE.md configurations many developers have built are probably over-engineered. Cherny's own CLAUDE.md is reportedly just two lines: one for auto-merging PRs, another for posting links to an internal Slack channel. That's it. He recommends deleting yours and starting over with each new model release.
Agents all the way down
Subagents came up repeatedly. The architecture Cherny describes (and has shared previously on his X account) treats Claude Code instances as modular workers: one writes code, another simplifies it, another runs end-to-end verification. His code review setup spawns several subagents in parallel, with a second wave specifically tasked with poking holes in the first wave's findings. "In the end, the result is awesome," he's said elsewhere, "it finds all the real issues without the false ones."
The podcast discusses what the original source calls a "Mama Claude" pattern: a primary Claude Code instance that recursively spawns sub-instances for subtasks. Cherny has noted that some Anthropic engineers now spend over $1,000 per month on Claude Code credits, primarily on code migration work where a lead agent creates task lists and delegates to ten or more subagents running simultaneously.
There's a practical ceiling here that Cherny acknowledges indirectly. Subagents can't spawn their own subagents. And the "Plugins" feature reportedly built by an agent swarm over a weekend, according to the original source, is the kind of claim that's easy to make on a podcast and hard to verify. The line between "AI built this" and "AI built 80% of this while a human debugged the remaining 20% for three days" tends to blur in these conversations.
So what actually changes?
The most grounded claim in the interview is about role evolution, not elimination. Cherny describes a team where product managers, designers, and finance people all write code through Claude Code. "Every single function on our team codes," he told Business Insider. Engineers, meanwhile, are drifting toward writing specs and talking to users, tasks that used to belong to PMs.
This tracks with a broader pattern. Anthropic hires "mostly generalists" now, according to Cherny. The bimodal distribution he describes (deep specialists on one end, broad generalists on the other) suggests the middle of the engineering skill curve is getting compressed, not that engineers are disappearing.
His old boss Dario Amodei predicted in March 2025 that 90% of code would be AI-written within six months. That timeline came and went. A LessWrong analysis found the claim didn't hold up even within Anthropic itself, depending on how you define "code." Cherny's version of the prediction is more hedged: coding is solved for him, and he thinks it'll get there for everyone. The gap between those two statements is where most of the industry actually lives.
Andrej Karpathy, OpenAI's former head of AI, noted in January that his manual coding skills have started to "atrophy" from reliance on AI tools. If the person who popularized the concept of AI-assisted coding is losing his edge, that's worth paying attention to, though whether it's a feature or a bug depends on your time horizon.
The full episode covers TypeScript parallels, terminal design decisions, and why Cherny joined Anthropic after reaching principal engineer at Meta. It's available on YouTube and major podcast platforms.
