Specialized coordination turns swarm activity into useful progress
Source → Observed → Interpretation → Model implication
Agent swarms and the new model economics
View source →Cursor reports running its old and new agent swarms on the same task, with the same models and the same time budget, measuring the fraction of a held-out SQL test suite passed; the new swarm performed better in every tested model configuration.
Cursor reports that with Grok 4.5 the new swarm reached 80% of the held-out suite in four hours, while the old swarm was paused before its second hour after becoming unstable.
The new swarm uses planner agents to decompose goals and delegate work and worker agents to execute bounded pieces; Cursor reports that planners do not implement and workers do not plan so that each role can use its context for a narrower purpose.
Cursor reports that the old Grok 4.5 run accumulated more than 70,000 merge conflicts before it was paused, while the new run logged fewer than 1,000 over four hours.
Cursor describes additional differentiated coordination and verification mechanisms including reconciler agents for conflicting design decisions, neutral third-party agents for merge conflicts, outside agents for decomposing overgrown files, and stacked review agents using different review lenses.
Cursor reports that its earlier browser swarm peaked at roughly 1,000 commits per hour on Git, while the newer system can peak around 1,000 commits per second and therefore uses a purpose-built version control system for agent-scale coordination.
Across the tested SQLite runs, Cursor reports similar quality across model mixes but large cost differences: total reported costs ranged from $1,339 for an Opus 4.8 planner with Composer 2.5 workers to $10,565 for GPT-5.5 used for both planning and work.
This source provides a controlled comparison in which the models, task, and time budget are held constant while the surrounding swarm architecture changes. The redesigned environment produces better measured task outcomes and substantially less coordination failure, refining Selection by connecting context allocation, execution structure, verification, coordination, and model economics to system viability. It also supports Cooperation because the source compares two multi-agent configurations and reports that coordination mechanisms materially change whether concurrent work composes into useful progress. The persistent planner-worker separation, together with reconciler, merge, decomposition, and review functions, supports Specialization: differentiated capabilities cooperate as parts of a swarm whose system-level behavior and performance are the unit being engineered and evaluated.
REFINES. Refines the model by providing comparative evidence that agent-scale engineering fitness depends on the architecture surrounding the agents, while also supplying concrete support for Cooperation and Specialization. The source shows differentiated roles and coordination mechanisms operating as an integrated system, but it remains a controlled Cursor research experiment rather than evidence that the same architecture or advantage generalizes across ordinary production software-engineering environments.
What this does not establish
- It does not establish that the redesigned swarm architecture generalizes to ordinary production engineering teams or arbitrary software-engineering tasks.
- It does not establish that roughly 1,000 commits per second correspond to 1,000 useful changes per second; commit rate is system activity rather than equivalent engineering value.
- It does not establish operation of 1,000 concurrent coding agents; the source discusses hundreds of concurrent agents and reports commit throughput, but does not publish a concurrent-agent maximum for the new system.
- Cursor is evaluating its own swarm architecture, so the reported comparison is first-party evidence rather than independent replication.
- The SQLite experiment isolates the harness comparison better than observational production evidence, but it does not isolate the causal contribution of each individual coordination, context, verification, or specialization mechanism.
Across repeated production engineering work, which differentiated agent roles and coordination mechanisms preserve the cooperation advantage as task diversity, concurrency, and organizational constraints increase?