Agent Swarms Make Model Choice an Architecture Decision
Cursor's swarm experiments are less a case for more agents than a case for better boundaries, memory, and verification.
Cursor’s article about agent swarms is easy to read as a story about scale. I think the more useful lesson is quieter: swarms change model selection from a runtime setting into an architectural decision.
More Agents Do Not Create More Clarity
Parallelism magnifies coordination quality. If responsibilities are vague, more workers create faster ambiguity. If tasks have sharp boundaries and useful handoffs, workers can move independently without losing the larger intent.
When I build agent workflows, I want each step to leave behind an artifact that the next step can inspect: a task brief, a decision record, a patch, a test result, or a concise failure report.
The Real Unit Is a Verified Outcome
Cheap execution that creates rework is expensive. The useful unit is the cost of a completed, verified outcome: a feature that passes tests, a document workflow that extracts the right fields, or a research answer grounded in the right source material.
Review adds latency and cost, but it is often cheaper than letting a bad assumption propagate through a task tree. Reliable progress is the product.
Related: [[experience]], [[pelmen]].