The AI Coding Stack Nobody Drew on a Whiteboard: Cursor + Claude Code + Codex Are Now One Workflow

The most-shared piece in developer Twitter and Hacker News this week is The New Stack's "Cursor, Claude Code, and Codex are merging into one AI coding stack nobody planned." The headline is sensationalist; the underlying observation is precise. In the first week of April 2026, Cursor 3 (codenamed Glass) shipped a rebuilt Agents Window for orchestrating parallel agents, OpenAI published codex-plugin-cc to install directly inside Anthropic's Claude Code, and the early-adopter cohort started running all three together. By May, IBM Bob crossed 80,000 developers with a 45% productivity claim, and GitHub flipped Copilot to usage-based billing. We are firmly past "which AI tool wins" and into the era of layered tools.

1. The shape of the stack

What "composable" means in practice: Cursor sits on top as the editor and parallel-agent orchestrator; Claude Code runs the long-context architectural work in a side terminal; Codex (or the OpenAI plugin) handles repetitive transformation passes. None of these tools own the whole workflow, and developers are explicitly choosing not to standardise on one. The same split happened to the JS bundler stack in 2018 (webpack + Babel + ESLint + Prettier) and the data stack in 2021 (dbt + Airflow + Snowflake + Looker). It's a stable equilibrium, not a transition.

2. Why "70% of developers" is the wrong number to fixate on

Yes, more than 70% of developers report using AI coding assistants. But the real adoption metric is different: how many tools per developer? At the high-productivity end, the answer is two to three, not one. The teams pulling away run an autocomplete model in the editor, an agent-mode model for refactors, and a long-context model for design reviews — and they switch providers per task type. Lock-in to a single AI assistant is now the productivity bottleneck.

3. The economic shift — usage-based billing arrives

GitHub's May 3 move from per-seat to usage-based pricing for Copilot was the headline. The structural read is more interesting: when models compete on per-token cost-to-output ratio, vendors who can't show better context management or fewer retries get repriced fast. Anthropic's Claude family and OpenAI's Codex agents are converging on a model where you don't pay for the seat, you pay for the work, and the buyer is increasingly the engineering manager, not IT procurement.

4. What this means for outsourced agencies and freelancers

Three near-term implications. First, your hourly rate is now defensible only if your AI stack is at least as good as your client's in-house engineer's stack — clients can compare velocity. Second, document which tool you used for which part of the build; "the agent wrote the migration" is a fair deliverable note in 2026, not a confession. Third, when scoping a project, the question is no longer "how many engineer-hours" but "how many agent-hours plus how many review-hours" — and the second number is the one that determines quality.

5. The crossover with AEO and GEO

One overlooked ripple: AI coding tools are accelerating the AEO/GEO content-engineering work. The /ai-summary/ routes that AEO experts recommend, FAQ schema injection, structured-data JSON-LD generation — all of these are textbook tasks for Codex-style agents. The engineering team that adopts a composable AI stack also accidentally becomes the AEO/GEO team, because the marginal cost of producing AI-readable content drops to near zero. That's the secondary, quieter trend behind the headlines.

My Take

The early hype said one AI tool would eat the engineer's job. The 2026 reality is more interesting and more uncomfortable: a small number of engineers, each running a well-tuned stack of three or four AI tools, are eating the jobs of mid-sized engineering teams. The differentiator isn't the tools — anyone can subscribe. It's the orchestration: knowing which model to throw at which problem, when to break the task in two and parallelise, when to abandon the agent's draft. That orchestration skill is the new senior-engineer competence, and it is teachable. The teams that train it deliberately this year will be unrecognisably faster by Q4. The teams that wait for "the winning AI tool" will lose the year.