Coding AI: from assistant to self-improvement

Of all the things large language models turned out to be good at, writing software is the one the labs bet the business on. It began modestly, as autocomplete: GitHub’s Copilot and the Codex model behind it suggested the next few lines as you typed. What made coding special was that the work could be checked — code either runs or it doesn’t — and that gave the field a hard yardstick. SWE-bench, a test built from real bug reports in open-source projects, became the number everyone chased, and the chase pulled the tools from suggesting lines toward doing whole jobs.

The turn from assistant to agent came fast. Cognition’s Devin was billed as an AI software engineer; Anthropic’s Claude Code and OpenAI’s Codex let a model take a task and run it, with the human reviewing the result rather than typing it. This became a genuine rivalry and a real business: OpenAI’s Codex passed two million weekly users, Anthropic bought a software company outright to strengthen Claude Code, and by early 2026 Anthropic’s coding-driven revenue was closing on OpenAI’s. A startup economy grew alongside it — Cursor, Lovable and the scramble over Windsurf — priced on the same idea.

The wider public got a name for it — “vibe coding”, Andrej Karpathy’s phrase for prompting a model and shipping what comes back without reading it closely — and a run of cautionary tales: an agent that wiped a live database, maintainers drowning in machine-generated bug reports, and evidence that many agent-written fixes that pass the tests would still be rejected by a human reviewer.

The deeper reason the labs care sits underneath the revenue. A model that writes code can be pointed at the lab’s own work — DeepMind’s AlphaEvolve already improves Google’s systems from the inside — and Anthropic reported that most of the code it merges is now written by Claude. That is the thread’s open question, which Anthropic put plainly in an essay titled “When AI builds itself”: coding is the first place where AI measurably speeds up the building of the next AI. For now there are still humans in the loop at every step — but the loop is getting shorter.