Plox
AI Coding3 mins read

Cursor’s agent swarm test points to a cheaper way to scale AI coding

Cursor’s upgraded agent swarm rebuilt SQLite in Rust from documentation only, with planner agents directing cheaper worker models. The setup eventually reached 100 percent on the test suite in every configuration while the older swarm struggled with merge conflicts.

The Decoder

The Decoder

Cursor's agent swarm suggests cheaper models can handle most coding when frontier models plan the work

AI coding swarms illustration for Cursor’s planner and worker agent experiment
Image credits:Nano Banana Pro prompted by THE DECODER

The core test: rebuild SQLite in Rust from docs alone

Cursor compared an upgraded agent swarm with its predecessor by asking both systems to recreate SQLite in Rust using only documentation. Source code, internet access, the SQLite binary, and test suites were withheld. Every configuration of the new system eventually reached 100 percent on the test suite, while the older swarm became bogged down by merge conflicts. The result highlights the importance of coordinating agent work, not just running more agents in parallel.

Planner agents decide; cheaper worker models execute

Diagram showing planner nodes and worker nodes splitting context across an agent task tree
Image credits:Cursor

The upgraded swarm separates agents into two roles: planner agents use powerful frontier models to break goals into smaller tasks, while worker agents use faster, cheaper models to carry out the work. Cursor says this split helps manage context because planners do not write code and workers do not plan. The task tree adapts as the work progresses, reducing the chance that a single agent loses track of the overall goal during long jobs. For teams evaluating AI coding tools, the key takeaway is that architecture may matter as much as model strength.

The old swarm generated conflicts faster than progress

Line chart comparing merge conflicts in old and new Cursor agent swarm runs
Image credits:Cursor

Cursor’s earlier system used worker agents, a judge agent, and an integrator for conflicts, but the integrator became a bottleneck. The new swarm reached 1,000 commits per second, prompting Cursor to build its own version control system for agent-scale collaboration. In the Grok 4.5 runs, the old swarm produced 68,000 commits in two hours and accumulated more than 70,000 merge conflicts, while the new run stayed below 1,000. Cursor also used shared design documents, compile-time references, neutral conflict resolution, and multiple review perspectives to control drift.

Cost results favor cheaper workers, but planner quality still matters

Stacked bar chart comparing the cost of rebuilding SQLite by model mix
Image credits:Cursor

Costs ranged from $1,339 for the Opus hybrid to $10,565 for GPT-5.5 running alone, according to the article. Workers used at least 69 percent of tokens in every run and usually more than 90 percent, making worker pricing a major cost lever. In one comparison, GPT-5.5 workers alone cost $9,373, while the entire worker fleet in the Opus and Composer setup cost $411 at comparable quality. Cursor’s broader implication is clear: use frontier models where ambiguity is highest, then let cheaper models execute well-defined tasks.

Discover More