The ‘Coding Agent’ Is the Wrong Name. It’s an Everything Agent.

At Code with Claude, Anthropic’s developer event in London, an engineer named Jeremy Hadfield asked the room a question: who here has shipped a pull request in the last week that was written entirely by Claude? About half the hands went up. Then he asked the follow-up — who shipped one without reading the code at all? Nervous laughter. Most of the hands stayed up.

Anthropic now says most of its own software is written by Claude, including most of the code inside Claude Code itself. And it’s not just them. The same stretch, OpenAI was named a leader in enterprise coding agents by Gartner and is pushing its Codex agent into companies like Dell and Ramp. Two labs, one direction, no ambiguity: handing your code to an agent has gone from novelty to normal in about a year.

We keep calling these things “coding agents.” That name is going to make a lot of people miss what’s actually being built.

What’s actually happening

The framing on stage at Code with Claude wasn’t “AI helps you write code faster.” It was further out than that. Boris Cherny, who runs Claude Code, put it this way: “The default isn’t ‘I’m going to prompt Claude’ — the default is now ‘I’m going to have Claude prompt itself.'” Another engineer’s version: “getting out of Claude’s way. We like to say: let it cook.” Anthropic even demoed a feature where coding agents write notes to themselves, then consolidate those notes to get better at a codebase over time.

The companies showing up to demo weren’t just big engineering shops, though Spotify and Delivery Hero were there. The tell was the startups: Lovable, Base44, Monday.com, companies whose whole product is letting people build apps by describing them. Apps that help you build apps. The abstraction is climbing, and it’s climbing toward people who don’t write code.

One Anthropic lead said Claude is “probably as good as a midlevel engineer at writing code” today. Another said the end state they’re aiming at is Claude “basically being able to build itself.”

Why “coding agent” undersells it

Code is just the most legible thing to measure. It’s text, it either runs or it doesn’t, and you can count the pull requests. So that’s where the benchmarks point and that’s the name that stuck.

But strip it back and these aren’t code tools. They’re general agents that can hold a goal, break it into steps, use tools, check their own work, and try again until something works. Code happens to be the cleanest demonstration of that loop. The loop itself isn’t about code.

That’s the thing most of the coverage is missing. The headline story is “engineers write less code by hand.” The actual story is that the gap between a person with an idea and a working thing is collapsing. And code was only ever one expression of that gap.

When the demo is “Claude shipped a pull request,” developers argue about whether that’s good or bad for developers. Fair argument. But it buries the bigger one: the same capability that writes a pull request can build the small internal tool a marketer needs, the script that reconciles two messy spreadsheets, the pipeline that turns a folder of notes into a published draft. None of that is “software development” in the job-title sense. All of it is now buildable by someone who can describe what they want.

The people this actually changes

I’m not an engineer. I run content and operations. A year ago, if I needed a tool that didn’t exist, I had three options: buy something close enough, wait for someone to build it, or do without. Now there’s a fourth — describe it and have an agent build it, for me, around how I actually work.

That’s the unlock the “coding agent” frame hides. The biggest beneficiaries of agents that write code aren’t the people who already write code. They’re the people who never could. The leverage doesn’t accrue to engineers getting 20% faster. It accrues to operators going from zero tools to their own tools.

And it lines up with everything else shipping right now. The same week as Code with Claude, Google showed search agents that build you a custom mini-app from a single question. Different company, same vector: software is becoming something you ask for instead of something you hire for. The startups demoing “apps that build apps” are just the middle of that curve. The end of it is everybody.

The catch worth naming

This isn’t a clean win, and the room knew it. Outside the conference, plenty of developers are pushing back hard. Researchers warn the flood of unreviewed code creates security and maintenance debt nobody’s paying down yet. Some engineers say their own skills are atrophying. Even Anthropic’s people admitted the old software best practices still apply, and that a lot of teams have lost sight of them in the rush.

Take that seriously. Shipping code you can’t read, at scale, into production people depend on, is how you build a slow-motion disaster.

But notice who that warning is actually for. It’s for teams handing off oversight on systems where the blast radius is other people. For an operator building a small tool for their own workflow, the blast radius is their own afternoon. The downside is a script that doesn’t run. The upside is leverage that didn’t exist a year ago.

So call it what it is. Not a coding agent — an everything agent that happens to be best at code right now. The story isn’t that machines are writing software. It’s that the door to building software just opened for everyone who was locked out of it, and most of them haven’t walked through yet.