Agent swarms and the new model economics

An artificial intelligence research group said it has developed a new agent swarm architecture that outperforms its previous system on complex software engineering tasks while reducing compute costs. The group said it tested the new swarm against the old system by having both build a SQLite implementation in Rust from documentation alone, using identical models and time budgets. The new swarm achieved higher pass rates on a held-out SQL test suite across every model configuration tested. Using a Grok 4.5 model, the new swarm reached 80% pass rate in four hours, while the previous system spiraled and was paused before its second hour. The architecture separates planner agents, powered by advanced models, from worker agents, powered by faster and less expensive models. The group said varying model assignments produced similar quality results with vastly different costs. The system uses a tree-like task decomposition that the group said improves context efficiency by preventing planners from filling context with low-level details and workers from losing sight of narrow tasks. The group said it built a new version control system from scratch to handle peak commit rates of roughly 1,000 commits per second, up from 1,000 per hour in earlier experiments. Coordination mechanisms include shared design documents with compile-checked references, a neutral agent to resolve merge conflicts, and automated decomposition of bloated files. The group said it has used the system internally for vulnerability discovery, test coverage improvement, and synthetic training data generation.