Blog
The AI Coding Stack Is Composing Itself
---
Nobody sat down and designed the new AI coding stack. It just... happened.
If you look at what engineers are actually running in production right now, three tools keep showing up together: Cursor for the IDE experience, Claude Code for deep agentic refactoring, and OpenAI Codex for async task workflows. Each one is excellent on its own. But when you combine them — Cursor handling orchestration and the writing surface, Claude Code running the heavy-lifting refactors and multi-file reasoning, Codex chewing through background task queues — you get something nobody explicitly architected: a three-layer composable coding stack.
The New Stack called it a stack with orchestration, execution, and review layers. That framing is exactly right, even though nobody at Cursor or Anthropic or OpenAI set out to build a layer. Developers just started composing these tools because each one solves a different part of the problem well. The stack emerged from usage, not from a roadmap.
From "Write Code Faster" to "Move Software Through the System"
Here is the part that matters more than speed.
Last week, LangChain published a deep dive on what they call "agentic engineering" — multi-agent systems that mirror how real engineering teams actually work. Not a single smart assistant, but swarms of agents with defined roles, shared context, and a lightweight leadership layer. Worker agents plan, execute, validate. A leader agent coordinates, governs, and maintains memory for the swarm.
The results are eye-catching. Cisco engineers running a pilot of this architecture on LangGraph reported that coordinated agent swarms cut debug time by 93% and compressed cross-team delivery timelines. The key insight: the biggest step-change doesn't come from better tools alone. It comes from systems that mirror real-world teams.
That reframes the entire conversation. For the past two years, the question has been "how do we write code faster?" The agentic engineering crowd has moved past that. The new question is "how do we move software through the system faster and safely?" — from requirements to design, development, security, testing, deployment, and operations. Every stage is amenable to at least partial automation, and some stages are ready for full end-to-end orchestration when agents collaborate cross-functionally.
This is where the composable stack gets powerful. Cursor Automations — launched March 5th — already lets you set agents to run triggered by new commits, incoming Slack messages, or scheduled timers. Replit Agent 4, shipped March 11th, added parallel task forking that auto-resolves merge conflicts roughly 90% of the time. These are not autocomplete features. These are background workers in a pipeline.
The Adoption Numbers Are No Longer Theoretical
Cloudflare just published their internal AI engineering data from the last 30 days, and the scale is remarkable:
3,683 internal users actively using AI coding tools — that's 93% of their R&D organization and 60% company-wide
47.95 million AI requests processed
241.37 billion tokens routed through AI Gateway
295 teams utilizing agentic AI tools and coding assistants
Their own assessment: "We've never seen a quarter-to-quarter increase in merge requests to this degree." When 93% of your engineering org is running AI tools daily and your merge request volume jumps to unprecedented levels, the "will developers adopt this?" question is settled. They adopted it. The question now is what happens to the discipline when the tooling becomes the default.
Business Insider reported last week that the transformation has already moved past engineering. Product managers and designers at startups like Speak are using Claude Code to write code and open pull requests. Engineers are taking on more product and design responsibility. Andrew Hsu, Speak's CTO, frames it plainly: specialization becomes less important when you have models that let you learn anything quickly. The winning profile is someone who can move up the ladder of abstraction and take on more agency in what they do.
OpenAI's own jobs impact report, published last Thursday, mapped AI's near-term impact across over 900 occupations covering 99.7% of US employment. They found that 18% of jobs face relatively high AI exposure. Greg Brockman, OpenAI's cofounder, wrote on X that AI has dramatically accelerated software engineering and is "on track to bring this same transformation to every other kind of work that people do with a computer."
So What?
Three things are happening simultaneously, and each one reinforces the other:
The composable stack means you pick the best agent for each layer instead of waiting for one tool to do everything. Cursor for speed and UI, Claude Code for depth and reasoning, Codex for background throughput. The stack doesn't need a vendor to bless it. It composes itself from developer choices.
The multi-agent architecture means those layers can actually coordinate. When agents share context, maintain memory, and run through a lightweight leadership protocol, you stop getting faster autocomplete and start getting a system that can take an intent from requirements to deployed code with human checkpoints along the way.
The adoption curve means this is not a pilot program anymore. When Cloudflare is routing 241 billion tokens a month through their internal AI gateway and 93% of their R&D org is daily-active, the infrastructure question is answered. The pipeline is real, it is production-grade, and it scales.
The AI coding stack composed itself because engineers reached for the best tool at each stage and the tools turned out to fit together. The multi-agent layer is forming on top because coordinated swarms deliver compounding returns that isolated assistants cannot. And the adoption data says the entire industry is moving together — fast.
If you are still thinking about AI coding tools as "autocomplete on steroids," you are looking at the wrong layer. The real shift is happening one level up: in the orchestration between agents, in the shared memory across a swarm, and in the composability of a stack that nobody designed but everyone is already building.
---
Cover image generated with Gemini 3 Pro via Batjko Labs.