Cursor's swarm rebuilt SQLite from scratch. Same scores, wildly different bills

Cursor's agent swarm implemented the 835-page SQLite manual in Rust from a blank page, with no source code, no test suites, no binary and no internet. It was graded on sqllogictest, the project's suite of millions of queries with known answers, and never told the suite existed. All four configurations of the redesigned harness eventually validated 100% of it.

The experiment compares the old and new harnesses on the same task, with the same models and the same time budget. The new one beat the old one in every configuration. With Grok 4.5 in both roles it hit 80% in four hours, while the old swarm stalled and was paused before its second hour ended. Cursor warns that trends matter more than single readings: some agents score low for hours before a late surge, others peak early and plateau.

The coordination tax: 70,000 conflicts

The activity data reads like productivity until you set it next to the results, the same blind spot that leaves most enterprises unable to say what their AI coding agents actually did. The old Grok 4.5 run made 68,000 commits in its first two hours, about 70 times the new run's pace, and piled up more than 70,000 merge conflicts, accelerating instead of settling. The new run logged under a thousand conflicts in four hours. The old run's most contested file drew 7,771 conflicts from 1,173 different agents. The new run's hottest file saw 47.

The final code shows the same gap. The old run sprawled across 54 Rust crates, including three separate SQL packages, the signature of two planners quietly building the same thing in different corners of the codebase. Cursor calls this split-brain, the coordination problem at the heart of parallel agent orchestration. The new run settled on nine crates and never added another. In the Fable 5 mix the old harness needed 64,305 lines of engine code to pass the suite; the new one did it in 9,908. In the Opus mix: 19,013 lines at 97%, versus 4,645 at 100%.

The pace forces the engineering. The earlier Browser swarm peaked near 1,000 commits per hour on Git; the new system peaks near 1,000 per second on a version control system Cursor built from scratch.

Most of the fixes around it are process, not model intelligence, and the process layer is where agents in production tend to fail. A neutral third agent resolves merge conflicts, the way a human merge queue would, just as Qoder gives its swarm a single coordinating leader instead of peer-to-peer messaging. Workers can flag oversized megafiles, blocking new commits until another agent splits the file. Design decisions go into shared documents that code references at compile time; when two planners contradict each other, a reconciliation merges the documents and the references propagate the resolution downstream. Agents may even make deliberate breaking changes outside their remit if they leave a comment, and every agent that hits the fallout reads the reasoning. One failure mode Cursor names is ossification: agents had learned not to touch critical code even when it needed to change.

Where the bills diverge

The economic finding is the one worth stealing. Every model split produced similar quality, Cursor says, while costs varied enormously, the kind of spread most enterprises cannot even see, per this report on invisible agent spend. The run using GPT-5.5 for both roles billed $9,373 in workers alone. A solo GPT-5.5 run cost $1,339. At four hours the new runs sat between 73% and 85%, the old between 11% and 77%, and every new configuration eventually finished at 100%.

Workers consumed at least 69% of tokens in every run, more than 90% in most, but planner tokens cost more. In the Opus 4.8 and Composer 2.5 mix, Opus produced a small fraction of the tokens and about two-thirds of the cost. Model choice alone does not set the bill: the Fable 5 planner invoiced slightly less than the Opus 4.8 planner despite a per-token price roughly twice as high, because it used far fewer planning tokens. Then its workers consumed several times more tokens, and the run cost markedly more.

Few steps in a large task need frontier intelligence: the initial decomposition, design decisions, certain arbitrations. Once a top planner has converted ambiguity into an explicit instruction, cheaper models only need to execute it. The same instinct shows up elsewhere in agent engineering, including a model router that picks the cheapest model for each task. A routing system called CodeRescue uses execution feedback to decide when a cheap model should retry, reaching near-perfect resolution rates at 35% of the cost of systematic escalation.

How the four configurations stacked up, as reported by Cursor:

The spec becomes the unit of work

Each jump in model capability has raised the level an engineer works at, Cursor argues. Autocomplete moved engineers from line to line, early models to code blocks, agents to files and features. With swarms, the unit of work becomes the specification: 835 pages of prose in, a database out. The rare part was stating the intention well enough. The swarm starts to resemble a compiler, decomposing a goal into a tree of tasks and converting it step by step into executable work; the matching lesson from Microsoft's MagenticLite research is that orchestration, not raw model size, carries complex agentic tasks. A compiler preserves meaning at every step; the swarm stays probabilistic at every step, and most of the machinery Cursor describes exists to close that gap. Ronald Coase would recognize the shape: coordination costs grow faster than the work itself, so organizations form bounded layers instead of everyone talking to everyone.

Two mechanisms stand out. The Field Guide is a folder owned by the agents, its index.md injected into every agent at startup under a line budget. Model weights are frozen, the reasoning goes, so unexpected situations are exactly what deserves recording. Cursor also tested review angles from a worker's full transcript to nothing but the codebase, with reviewers on different models, training runs and personalities. No single angle catches everything, but decorrelated angles miss less, the way self-driving systems beat humans in aggregate without any perfect component.

A footnote shows how fragile the stack still is. Cursor wanted GPT-5.6 Sol for its flagship configuration, but the new model spiraled out of control on literal, insistent phrasing, so it reverted to GPT-5.5 rather than tune one model and skew the comparison. The harness's gains sit on top of model behavior that is still uneven.

The output of the Opus 4.8 solo run is public at github.com/cursor/minisqlite for anyone to take apart; Cursor says it looks excellent at first glance and has not done a deeper manual review. The bottom line: the harness went from barely working to reliably working, and that reliability turned the model mix into a cost decision instead of a quality gamble.