Cursor's New Agent Swarm Architecture Splits Planning and Coding to Cut Costs by Up to 15x

Cursor says its redesigned multi-agent coding system proves that cheap language models can handle the bulk of software engineering work, as long as a more capable model handles planning. The company detailed the architecture, which ships as part of Cursor 3, in a benchmark comparing it against an earlier version of its agent swarm.

In the test, both the old and new swarms were asked to rebuild SQLite in Rust using only the 835-page SQLite manual — no source code, no test suite, no internet access. Success was measured against sqllogictest, a suite of millions of SQL queries with known answers that the swarms were not told about in advance.

Every configuration of the new system eventually reached 100 percent accuracy on the hidden test suite. The old swarm scored between 11 and 77 percent after four hours and got bogged down resolving its own merge conflicts, according to Cursor.

Planners decide, workers execute

Cursor's new architecture assigns two roles. Planner agents, running frontier models, recursively break a large goal into a tree of smaller tasks. Worker agents, running faster and cheaper models, execute those tasks without doing any planning themselves. Cursor says this split addresses a context problem: a single agent working alone has to hold both the overall goal and the current task in mind while traversing an entire task tree, which the company says explains why long-running agents tend to drift off course.

Cursor tested four configurations: GPT-5.5 running solo, Grok 4.5 running solo, Opus 4.8 as planner with Composer 2.5 as worker, and Fable 5 as planner with Composer 2.5 as worker. All four beat the equivalent old-swarm configuration.

The difference in efficiency was stark. In the Grok 4.5 comparison, the old swarm produced 68,000 commits in two hours — about 70 times more than the new swarm — but accumulated more than 70,000 merge conflicts, versus fewer than 1,000 for the new system. The old run split the project into 54 Rust crates; the new run settled on nine. In the Opus configuration, the old system needed 19,013 lines of code to reach 97 percent accuracy, while the new architecture reached 100 percent with 4,645 lines — an 85 percent reduction at a higher score.

Cheap workers drive most of the cost savings

Total run costs ranged from $1,339 (Opus planner, Composer 2.5 worker) to $10,565 (GPT-5.5 solo) — a roughly 15x spread for comparable results. Workers accounted for at least 69 percent of tokens generated in every run, and often more than 90 percent, making worker pricing the dominant cost factor.

Composer 2.5, the worker model Cursor used in its cheapest configuration, is priced at $0.50 per million input tokens and $2.50 per million output tokens. According to Cursor founder Michael Truell, Composer 2.5 is built on Kimi K2.5 and benchmarks at a level comparable to Opus 4.7 and GPT-5.5 — a claim from Cursor that has not been independently verified against standard public benchmarks. In the GPT-5.5 solo run, workers alone cost $9,373; the equivalent worker fleet running Composer 2.5 cost $411 for comparable quality, according to Cursor's own figures.

To handle coordination at scale, Cursor said it built a custom version control system after an earlier browser-based swarm reached roughly 1,000 commits per hour and its integrator agent became a bottleneck. The new system reached 1,000 commits per second. Cursor also described a "split-brain" failure mode, where planners unaware of each other would independently build duplicate functionality in different ways, and a "field guide" — a size-capped knowledge document maintained by the agents themselves and loaded at the start of every new agent session.

Cursor published the codebase from the Opus solo run as an open-source project called minisqlite on GitHub. The company noted that a prerelease version of Fable 5 was separately used to help rewrite the Bun JavaScript runtime from Zig to Rust, with 64 agent instances producing more than a million lines of code over 11 days for about $165,000.

What this means

These are Cursor's own benchmarks, run on a synthetic, closed-book test rather than a production codebase, and the company has commercial incentive to show its architecture off favorably. Still, the core claim — that separating planning from execution lets cheap models do most of the work — lines up with a broader industry pattern of routing tasks to the smallest model that can handle them. The bigger caveat is deployment reality: a late-2025 study cited in Cursor's own writeup found that 68 percent of production AI agents stall after ten steps or fewer, meaning today's economics gains largely apply to controlled experiments, not yet to typical day-to-day agent use.