Everyone assumed AI coding tools would eventually merge into one. Instead, the opposite happened.
What happened in the first week of April 2026
Three things, all within one week:
- Cursor v3 (codenamed Glass) launched Agents Window, which can simultaneously manage multiple AI Agents across local machines, Git worktrees, and cloud sandboxes.
- OpenAI released
codex-plugin-cc— an Apache 2.0 open-source plugin that runs directly inside Claude Code, with six slash commands. - Early adopters began chaining all three tools together. Nobody planned this path, but it formed naturally.
This is not any single company's strategy. It is the result of the ecosystem spontaneously layering itself.
What the three-layer collaboration stack looks like
If you are one of those early adopters, your workflow now probably looks like this:
Layer 1 — Orchestration: Cursor
No longer "looking at code in an editor." Instead, you open several Agents in the Agents Window, assign tasks, and watch each Agent's progress. Design Mode lets you mark UI issues directly in the built-in browser and feed visual feedback to the Agent. You are managing a small team, not typing.
Layer 2 — Execution: Claude Code + Codex
These two tools compete and cooperate at this layer. Claude handles tasks that require context understanding; Codex handles tasks that can run in parallel. The /best-of-n command sends the same task to multiple models simultaneously, runs them in isolated worktrees, and compares the results.
Layer 3 — Review: Codex reviews Claude
/codex:adversarial-review — Codex performs an adversarial review of code written by Claude, specifically checking for authentication vulnerabilities, data loss risks, and race conditions. When issues are found, a Gate can be set to automatically block merges.
"Having the same model that wrote the code review it is like letting a student grade their own homework."
Cross-vendor independent review solves this problem.
Why OpenAI put Codex inside Claude Code
This is the most interesting move in the whole story.
OpenAI did not try to stop users from also using Anthropic's Claude Code. Instead, it proactively placed a plugin inside it, making it easier for Codex to be invoked within Claude Code's workflow.
The logic: usage-based billing, distribution over lock-in.
Codex charges by API usage. Users may use Claude Code, but if they call Codex inside it for adversarial review or parallel tasks, OpenAI still gets paid. Getting into Claude Code's user pool is easier to achieve than hoping to pull users away from Claude.
This is the new model for AI tool competition in 2026: I don't eliminate you; I infiltrate your user base and collect revenue.
The numbers now
- Claude Code received a 46% "most liked" rating in a survey of 906 engineers.
- Codex surpassed 3 million weekly active users (it was 2 million a month ago).
- Claude Code accounts for approximately 4% of public GitHub commits (March 2026).
- Claude Code's annualized revenue is estimated by analysts to exceed $2.5 billion.
The market is not winner-takes-all; all three tools are growing.
What this means
The evolution path of AI coding tools is different from SaaS. In the CRM world, the endgame was Salesforce dominating everything. In version control, the endgame was GitHub monopolizing.
AI coding tools may not follow that path. The reason is that each layer requires different expertise, and composition is more efficient than monopoly:
- Orchestration needs Cursor's multi-Agent UI.
- Code generation needs top-tier reasoning models.
- Cross-validation needs an independent perspective.
The three tools are not in a competitive relationship; they are in a division-of-labor relationship. Developers act as their own integrators in the middle.
How far this path can go is still unknown. But what happened in the first week of April shows that both Anthropic and OpenAI assume developers will use multiple tools, not pick one. That itself is a significant change.
Sources: Cursor, CocoLoop, Claude Code, and Codex are merging into one AI coding stack nobody planned (The New Stack)