Your AI Coding Tools Aren't Competing Anymore. They're Stacking.

In one wild week in April 2026, the AI coding tool market stopped trying to pick a winner and started building a layered stack nobody designed on purpose.

There was supposed to be a winner. That was the story we all kept hearing. The AI coding tool market would eventually consolidate, developers would standardize on one product, and everyone would move on. Neat and tidy. Except, that's not what happened. Not even close.

In the first week of April 2026, three things landed almost simultaneously, and together they told a very different story. Cursor shipped a completely rebuilt interface for running multiple AI agents at once. OpenAI published an official plugin that runs Codex directly inside Anthropic's Claude Code. And early adopters started running all three together, not as alternatives, not as competitors, but as layers in a stack. A stack nobody sat down to design, but that's assembling itself anyway.

Sound familiar? It should.

If you've been around infrastructure long enough, you've seen this pattern before. Nobody runs a single observability tool. You run Prometheus for metrics, Grafana for dashboards, and PagerDuty for alerts. Each does one thing well. The value isn't in any single tool, it's in how they talk to each other.

AI coding tools are doing exactly the same thing. Instead of collapsing into one product, they're splitting into specialized layers. And honestly, that makes a lot more sense than one tool trying to do everything.

Three launches. One very clear signal.

Let's break down what actually dropped. First, Cursor 3 (codenamed Glass) replaced its old Composer pane with something called the Agents Window, a dedicated control surface for managing fleets of AI agents running in parallel. Local machines, cloud sandboxes, worktrees - all in one sidebar. There's a /best-of-n command that sends the same prompt to multiple models simultaneously so you can compare outputs. It's less of an editor with AI bolted on, and more of a control plane for orchestrating AI work at scale.

Three days before that, OpenAI quietly published a plugin on GitHub called codex-plugin-cc. It installs inside Claude Code - Anthropic's terminal-based coding agent and gives you six slash commands. You can run a standard code review,  trigger an adversarial review that pressure-tests edge cases around authority and race conditions, or hand a task off to Codex entirely as a subagent. There's even a review gate that can block Claude's output until Codex signs off on it.

OpenAI just shipped an official integration into a direct competitor's product. That's not something that happens in a normal market.

Think about that for a second. OpenAI built a plugin for Anthropic's product. Apache 2.0 licensed. No walled garden. No new runtime required. Just Codex, invoked from inside Claude Code, using your existing auth. That's not the conventional playbook. That's something else entirely.

Three layers are forming.

What's emerging looks less like a product choice and more like a toolchain with distinct responsibilities. Here's how the stack is shaking out:

  • Orchestration: Cursor 3, Antigravity. Manages agents, coordinates parallel work, provides the control surface.
  • Execution: Claude Code, Codex. Actually writes, reviews, and debugs the code. Lives in the terminal.
  • Review: Cross-provider verification. One model writes, another challenges. Independent scrutiny.

The review layer is the newest and, frankly, the most interesting. There's an obvious problem with asking one model to review code it just wrote. That's grading your own homework. The structural bias is unavoidable, the model shares its own assumptions, blind spots, and optimization quirks. A second model from a different provider, trained differently, catches entirely different classes of problems. That's not a marketing claim. That's just how it works.

Why is OpenAI playing this differently than expected?

Here's the part that surprised a lot of people. The conventional wisdom says: build a walled garden, make switching painful, lock users in. OpenAI looked at Claude Code's massive and enthusiastic developer base and said, what if we just went to where the developers already are?

Every Codex review triggered through the plugin counts against the developer's ChatGPT subscription or API key. Zero acquisition cost. Incremental billing. Anthropic gets a richer plugin ecosystem. OpenAI gets distribution inside a competitor's installed base. Nobody loses.

Both companies recognized developers will use multiple tools regardless. The question is whether your tool is in the stack or outside it.

This is pragmatism, not altruism. And it's a signal that the "pick one tool and stick with it" era for AI coding might already be over.

What this means if you're actually building software.

A few things change if this composable pattern holds. Model selection starts to look like infrastructure decisions, you pick Claude for nuanced reasoning across long contexts, Codex for throughput on parallelizable tasks. The code editor, which has been the center of gravity in software development for 40 years, starts to recede. Cursor's Agents Window is already competing with the editor as the primary interface. And code review shifts toward being adversarial by default, cross-provider review catching the things single-model workflows will always miss.

None of this is guaranteed to stabilize cleanly. More layers mean more complexity, more costs stacking up, and steeper learning curves for developers who aren't power users. GitHub Copilot has its own agent capabilities in the works. AWS Kiro just shipped an agent-first IDE. Every major cloud provider now has a position in this space. The next phase will be determined by which layers become commodities and which become the new control points.

But the narrative that one tool would win? That's probably done. The AI coding market isn't consolidating. It's composed. And for developers who already know how to think in toolchains, that's not a complication - it's a feature.