The Big Idea: Harness Over Model
Cursor, the AI-assisted coding platform, has been sending a clear signal over the past two months: its future lies not in the IDE, but in the orchestration layer that sits around AI models. On Wednesday, the company shipped the Cursor SDK; the next day, its harness team published a detailed explainer on the agent harness. Together, these releases package years of internal orchestration work and make it available to any developer. In February, CEO Michael Truell published an essay declaring this the “third era” of AI software development. And just last week, Cursor announced a partnership with SpaceX to train its proprietary Composer models on xAI’s Colossus supercomputer.
Read together, the message is unmistakable: Cursor believes the AI model itself is becoming a commodity, and the product that will win the next decade is the harness around it. The strongest confirmation came from outside Cursor: Google told The New Stack this week that it doesn’t care which coding tool developers use—Gemini, Claude Code, or Cursor—reinforcing the idea that the model is interchangeable.
From IDE to Agent Platform: Cursor’s Evolution
For months, Cursor has been shedding its identity as just an IDE company. When Cursor 3 shipped earlier this month, Jani MSV wrote on The New Stack that the IDE is now a fallback. Instead, agents spin up dedicated cloud VMs, work for hours, and return logs, video recordings, and live previews. According to Truell’s “third era” post, agent usage at Cursor has grown more than 15x in the last year.
Twelve months ago, Cursor had 2.5 times as many Tab autocomplete users as agent users. Today, it has 2x as many agent users as Tab users. Inside Cursor’s own engineering team, more than a third of internal pull requests are created by agents working in cloud VMs. Truell expects “the vast majority” of development work to look that way within a year.
The Commoditization of AI Models
Cursor’s bet hinges on a core thesis: AI models—whether from OpenAI, Anthropic, Google, or open-source providers—are becoming interchangeable commodities. The real differentiator is the harness that orchestrates them: the tooling that manages codebase indexing, agent coordination, observability, and deployment. Google’s recent statement—that it doesn’t care which coding tool developers use—bolsters this view. When models are fungible, the value moves to the platform that makes them easiest to use.
This is why Cursor is racing to position itself as an agent platform, not just an editor. The Cursor SDK is the key piece of that strategy.
Inside Cursor’s Agent Harness: The Cursor SDK
On April 29, Cursor released the Cursor SDK in public beta. It’s a TypeScript package (installable via npm install @cursor/sdk) that lets developers build agents directly on Cursor’s harness, model-agnostic, deployable locally or on Cursor Cloud against dedicated VMs. The harness ships with several core features:
- Codebase indexing for context-aware agent actions
- MCP server support for modular tool integration
- Subagents for parallel task execution
- Observability hooks for monitoring agent behavior
This puts Cursor in direct competition with OpenAI’s Agents SDK and Anthropic’s Claude Agent S. But unlike those offerings, Cursor’s harness is built on the experience of running millions of agent sessions inside its own IDE.
Agent Usage Surge: The Data Behind the Shift
The numbers underscore the trend. As noted, agent usage has grown 15x year-over-year. The user base has flipped: previously dominated by Tab autocomplete, now agents dominate. Cursor’s internal stats show that over a third of internal pull requests are agent-generated, running in cloud VMs. Truell predicts that within a year, the vast majority of development work will follow that pattern—meaning the IDE will matter less, and the harness will matter more.
This shift is not just about convenience; it’s about productivity. Agents can work asynchronously, spinning up isolated environments, running tests, and generating code while developers focus on higher-level design. The harness provides the reliability and observability needed to trust these agents.
What This Means for Developers and the Industry
If Truell is right—and many industry observers believe he is—the next decade of software development will be defined by agent orchestration, not model selection. Developers will choose the harness that gives them the best control, observability, and flexibility. Cursor is betting that its early lead in agent usage and its open SDK will make it the default platform for building AI-powered development workflows.
Meanwhile, companies like Google, OpenAI, and Anthropic will continue to improve their models, but the value capture may shift to the platforms that sit on top. Cursor’s SDK and harness are designed to be model-agnostic, allowing developers to swap models without changing their workflow.
For developers, the message is clear: learning to build with agent harnesses—whether Cursor’s, OpenAI’s, or Anthropic’s—will be a critical skill. The future of coding is not about typing code manually, but about directing AI agents to write it for you, with the harness providing the guardrails and infrastructure.
A New Era of Software Development
Cursor’s $60 billion valuation bet is not on a better model—it’s on the idea that the harness is the new platform. By releasing the SDK and doubling down on agents, Cursor is positioning itself at the center of the next wave of AI-assisted development. Whether you use Cursor, Claude Code, or Gemini, the underlying trend is the same: the model is a commodity, but the harness is where the value lies.