Cursor: AI Agent Swarm Builds 80% of SQLite in 4 Hours — at a Fraction of the Old Swarm's Cost

Cursor compared the old and new versions of its AI agent swarm on a single task: build the SQLite database engine from scratch in Rust, with only its documentation as input. The new swarm passed 80% of a hidden SQL test suite in four hours on Grok 4.5, while the old swarm stalled and was stopped before the two-hour mark.

What the experiment showed

The new Cursor swarm beat the old one in every model configuration, with the same time budget and the same models. The team measured the share of a held-out SQL test suite that each swarm passed, and by that metric the advantage was consistent rather than a one-off.

• The task: build SQLite from scratch in Rust using only the documentation
• Grok 4.5 in the new swarm: 80% of tests in 4 hours
• The old swarm on the same task stalled and was stopped before the two-hour mark
• The flagship of earlier experiments (early 2026) was a swarm that built a web browser from scratch: a proof of concept, not a polished product
• Inside Cursor, the swarm has already found and fixed vulnerabilities in open-source software, raised test coverage and generated billions of tokens of synthetic training data

Why does a swarm beat a single agent?

The swarm wins on context efficiency, not on parallelism as such, Cursor insists. Large tasks naturally decompose into a tree: the goal sits at the root and is recursively split into elementary units of work. Cursor's swarm is built around two roles. Planner agents, running on the smartest models, split the goal and delegate pieces. Worker agents, running on fast, cheap models, execute those pieces.

The key is that the planner never writes code and the worker never plans. As a result, the planner's context isn't clogged with low-level details, and the worker spends its entire context on one narrow task. A single agent, by contrast, has to hold both the overall goal and its current position in the tree in its head, and over time it drifts. The swarm's design is also a superset of more rigid orchestration schemes: instead of a fixed topology, its shape grows to fit the contours of the task. That is why the approach generalizes to building a browser, solving math problems and optimizing GPU kernels.

"We believe the ability to scale a swarm of agents stems from this context efficiency, not from parallelism as such," the

Cursor team wrote on the company blog.

What it costs

The cost of the runs varied enormously while quality stayed almost the same. This is the "new economics of models" from the study's title. Cursor varied the allocation of roles: in some runs a single model did everything, in others a frontier model planned while a fast, cheap model did the work. The final quality came out similar, but the bill was radically different. The economist Ronald Coase described similar logic back in 1937: coordination costs grow faster than the work itself, so organizations split into tiers of bounded units rather than letting everyone talk to everyone.

Version control is a separate engineering problem. Git and Cargo rely on coarse locks, which is fine for a single developer but breaks down with hundreds of parallel agents. According to Cursor, last year's browser swarm hit a ceiling of about 1000 commits per hour on Git; the new system sustains about 1000 commits per second, which is why Cursor built its own version control system.

What this means

Multi-agent systems are moving from "just launch more agents" to deliberate coordination engineering: the right division of roles between an expensive model and a cheap one delivers the same quality at a fraction of the cost. The economics now depend not only on which model to pick, but on how to distribute work among them.

Frequently asked questions

What is an AI agent swarm?

It is a system of many agents that jointly solve one large task. In Cursor's swarm, Planner agents (smart models that plan and delegate) work alongside Worker agents (fast, cheap models that execute), and the swarm's shape grows to fit the contours of the task.

Why build SQLite from scratch?

SQLite is a complex, well-documented database engine, which makes it a convenient test of the limits of agent coordination. Cursor gave the swarm only the documentation and measured the share of SQL tests passed; the new swarm on Grok 4.5 reached 80% in 4 hours.