The New Bottleneck

AI coding tools have outgrown their original job description.

One year ago, the typical AI coding assistant was a souped-up autocomplete. Type a function name, and the tool would predict the next few lines. It was useful in the way that spell-check is useful: a convenience that shaved minutes off familiar tasks without fundamentally changing the nature of the work.

That description no longer applies to any of the major players. In the space of a few months, the leading code editors have undergone a transformation so thorough that the category itself needs a new name. What used to be an “AI coding assistant” is now something closer to a junior developer, a project manager, or an entire prototyping department, depending on the product in question.

This shift matters even if one never intends to write a line of code, because the tools that reshape software development tend to affect everything else shortly afterward.

The Big Three (and a Newcomer)

The competitive landscape has consolidated around a handful of products, each representing a slightly different philosophy about what an AI-powered editor should be.

Cursor, built by the startup Anysphere, released its 2.0 update in late October and has been iterating aggressively since then. The headline feature was Composer, a proprietary coding model that completes most tasks in under 30 seconds, roughly four times faster than comparable alternatives. But the real change was architectural: Cursor 2.0 redesigned its entire interface around the concept of agents rather than files. Up to eight AI agents can now be run simultaneously on a single codebase, each working in an isolated environment through Git worktrees, while a developer monitors their progress the way a manager might review pull requests from a team. In February, Cursor added long-running agents that can plan and execute work autonomously over extended periods, and introduced subagents that spawn their own subagents, creating branching trees of coordinated work.

Windsurf, built by the AI company Cognition, has taken a similar multi-agent approach. Its Wave 13 update in December introduced parallel agent sessions, Git worktree integration, and a free tier for its SWE-1.5 model, a near-frontier coding AI that Cognition made available to all users at no cost for three months. Windsurf has also been rapidly integrating new models as they appear: in the first three weeks of February alone, the platform added support for Claude Sonnet 4.6, GLM-5, Minimax M2.5, and GPT-5.3-Codex-Spark. The pace of model additions suggests that the editor layer is becoming model-agnostic, a universal interface through which developers access whichever AI happens to be best at a given task.

Google Antigravity, announced alongside Gemini 3 in November, is the newest and most ambitious entrant. Built partly from talent and technology Google acquired when it bought the Windsurf team for $2.4 billion in mid-2025, Antigravity introduces what it calls a “Manager View,” a dedicated interface for the orchestration of multiple agents working asynchronously across different workspaces. Where Cursor and Windsurf still feel like code editors with agent capabilities bolted on, Antigravity’s design suggests that Google sees the future of development as something closer to project management: agents are dispatched to handle tasks, their deliverables (which Antigravity calls “Artifacts”) are reviewed, and feedback is provided in much the same way as one might comment on a shared document. It is free during its public preview, which puts considerable pricing pressure on the entire category.

And then there is Claude Code, Anthropic’s terminal-based coding agent, which operates outside the IDE paradigm entirely. Rather than sitting inside an editor, Claude Code reads an entire repository, understands the architecture, plans multi-step changes across multiple files, and delivers pull-request-ready code. By February 2026, its annualized revenue had reached $2.5 billion, having doubled since January 1. A UC San Diego and Cornell University survey from January found it neck-and-neck with GitHub Copilot and Cursor as one of the three most widely adopted platforms among professional developers. Perhaps the most telling data point: Microsoft, which owns GitHub and sells Copilot, has reportedly adopted Claude Code internally across major engineering teams.

From Autocomplete to Autonomy

The common thread across all of these tools is a shift from suggestion to execution. Last year, AI coding assistants offered predictions in response to the question “what should I write next?”. Now, it merely needs to be told “build this feature” in order to plan its approach, write the code across however many files are needed, run the tests, open a browser to verify the result, and present a summary of what it did.

This is not a subtle distinction. It changes the developer’s role from author to reviewer, from someone who writes code to someone who evaluates and directs it. The 2025 DORA report, Google’s annual survey of software development practices, found that 90% of developers now use AI daily, typically spending a median of two hours per day working with these tools. But the report also surfaced an important paradox: while individual developers report significant productivity gains, those gains do not automatically translate into better organizational outcomes. Teams using AI write more code faster, but code review time has increased by roughly 91% as the volume of AI-generated pull requests overwhelms human capacity to review them.

In other words, the bottleneck has moved. Writing code is no longer the hard part. The difficulty is now concentrated around knowing whether the code is good, whether it does what it should, and whether it introduces subtle bugs or security vulnerabilities. The tools that win this market will be the ones that solve the review problem, not just the generation problem.

The Vibe Coding Layer

An entirely separate category has also matured alongside the professional developer tools, this one aimed at people who do not know how to program at all. The practice now has a name: “vibe coding,” coined by computer scientist Andrej Karpathy and recognized by MIT Technology Review as one of its 10 Breakthrough Technologies of 2026.

Platforms like Replit, Lovable, Bolt.new, and Vercel’s v0 allow users to describe an application in plain English and receive a working prototype, sometimes within minutes. The quality varies, of course, and the gap between “impressive demo” and “production-ready application” remains substantial. Databases need to be configured, authentication systems need to be connected, and deployment requires decisions that most non-technical users are not even remotely equipped to make. Yet the trajectory is obvious: Y Combinator’s Winter 2025 batch included startups whose codebases were more than 90% AI-generated.

What this means in practice is that building a functional internal tool, a prototype, or a personal project no longer requires hiring a developer or learning to code. What it requires instead is the ability to describe what one wants clearly enough for an AI to build it, which is a skill closer to writing a good brief than to writing good Python.

What Actually Changed

Stepping back from the individual product announcements, the structural shift is straightforward. Software development used to be a craft constrained by how fast humans could type and how much complexity they could hold in their heads. AI tools have largely removed the first constraint. The second, understanding what the code should do and whether it does it correctly, has consequently become more important, not less.

This creates a counterintuitive dynamic. The DORA report’s central finding was that AI acts as an amplifier: teams with strong engineering practices become significantly more productive, while teams with weak practices simply produce more buggy code at a faster rate. As one analysis put it, speed without stability is just accelerated chaos.

The competitive picture for the tools themselves is evolving week by week. Cursor, Windsurf, and Antigravity are converging on similar feature sets (parallel agents, multi-model support, plan-then-execute workflows) while competing fiercely on price, with Google’s free offering putting particular pressure on the paid tiers. Claude Code occupies a different niche, operating at the level of entire repositories rather than individual files, but its explosive growth suggests that many developers prefer to skip the IDE wrapper entirely.

The most honest assessment of where things stand is probably that no one has yet figured out the right interface for this new kind of work. Running eight parallel AI agents and reviewing their output is genuinely useful, but it is also genuinely different from anything developers have done before, and the toolmakers are still experimenting on the problem of how to make it feel natural rather than overwhelming. The editor that succeeds in solving that problem, that manages to make directing AI agents feel as intuitive as typing code once did, will likely define the category for years to come.

In the meantime, the old joke about software eating the world needs an update. Software is still eating the world, but it is increasingly writing itself in order to do so.