What happened
Cursor.com developed a new agent swarm architecture, achieving 80% completion on building SQLite from scratch in Rust with Grok 4.5 in four hours; their previous swarm failed within two. This architecture employs a "trees and leaves" decomposition, utilising frontier models for planning and less expensive models for execution, maintaining similar quality with varied costs. A custom version control system manages concurrency, processing 1,000 commits per second, and resolves conflicts like split-brain designs and merge contention through shared design docs and neutral third-party agents.
Why it matters
Agentic development workflows gain significant efficiency and scalability, shifting unit economics for complex software projects. This new architecture, by decomposing tasks into planner and worker roles, improves context management and allows cheaper models to execute work without quality degradation. The custom version control system removes a critical constraint on agent concurrency, enabling 1,000 commits per second and resolving coordination failures. Procurement teams and architects evaluating agentic solutions must now prioritise systems demonstrating proven context efficiency and high-throughput conflict resolution, moving beyond single-agent benchmarks, following Databricks' recent findings on cost-effective coding agents.
