Experiments testing agent cooperation revealed a new tier of task scale and complexity, exemplified by building a web browser from scratch. While a proof of concept, it highlighted the need for deliberate engineering.

Returning to building SQLite from documentation, a new swarm outperformed the old one using the same models and time. Using Grok 4.5, the new swarm achieved 80% accuracy in four hours, significantly surpassing the older swarm. Model mixing—where a frontier model plans and a faster model executes—showed similar quality but vastly different costs.

The swarm's design, mirroring trees, uses Planner agents (smart models) to subdivide tasks and Worker agents (faster models) to execute them. This flexible topology scales compute and context with task complexity, generalizing across diverse tasks. This decomposition also enhances agent performance through context efficiency, preventing drift by separating planning from execution.

A new version control system (VCS) handles the swarm's ~1,000 commits per second rate, enabling coordination mechanisms. Failure modes at this speed, such as split-brain design and contention, are addressed through prompting, shared design docs, impartial merge conflict resolution, flagging "megafiles," and "intentional breakage" for core code changes. Stacked review lenses, employing decorrelated approaches, ensure sustained quality.