---
title: "Agent swarms and the new model economics"
description: "Agent swarms are evaluated as a way to scale software work beyond what a single long-running agent can reliably manage. The test revisits building SQLite in Rust from its documentation, comparing old and new swarms with identical models and time budgets against a held-out SQL suite; the new system performed better in every configuration, reaching 80% with Grok 4.5 in four hours. Planners decompose task trees while workers execute narrower pieces, supported by a purpose-built version control system and mechanisms for design conflicts, merge collisions, megafiles, intentional breakage, and review. Reported results include a full-suite pass using 9,908 engine-code lines versus 64,305 for the old harness, while costs ranged from $1,339 for an Opus 4.8 hybrid to $10,565 for GPT-5.5 alone, indicating that cheaper models can perform most execution after frontier planning resolves ambiguity."
---

# Agent swarms and the new model economics

[Cursor](https://yomu.fyi/company/cursor) · Wilson Lin · Jul 20, 2026

**Type:** Benchmark

## Summary

Agent swarms are evaluated as a way to scale software work beyond what a single long-running agent can reliably manage. The test revisits building SQLite in Rust from its documentation, comparing old and new swarms with identical models and time budgets against a held-out SQL suite; the new system performed better in every configuration, reaching 80% with Grok 4.5 in four hours. Planners decompose task trees while workers execute narrower pieces, supported by a purpose-built version control system and mechanisms for design conflicts, merge collisions, megafiles, intentional breakage, and review. Reported results include a full-suite pass using 9,908 engine-code lines versus 64,305 for the old harness, while costs ranged from $1,339 for an Opus 4.8 hybrid to $10,565 for GPT-5.5 alone, indicating that cheaper models can perform most execution after frontier planning resolves ambiguity.

## Context

The experiments began from a hypothesis that cooperating agents could handle greater task scale and complexity. A previous browser-building swarm demonstrated the concept but produced software far short of polished quality, motivating a more deliberate evaluation against SQLite, a task the earlier swarm had struggled with.

## Approach / What changed

The system decomposes work into a tree managed by planner agents and worker agents, then uses a purpose-built version control system and specialized coordination mechanisms for design conflicts, merge collisions, megafiles, intentional breaking changes, and review. Evaluations compared old and new swarms with the same models and time budgets, while varying model assignments and measuring held-out SQL-suite performance and cost.

## Takeaways

- The new swarm outperformed the old swarm in every tested model configuration; with Grok 4.5, it reached 80% on the held-out SQL test suite in four hours, while the old swarm was paused before its second hour.
- Planner agents handle decomposition and design decisions, while worker agents execute narrower tasks. The post attributes the resulting scalability largely to context efficiency rather than parallelism alone.
- Workers produced at least 69% of tokens in every run and more than 90% in most, but planner pricing strongly affected total cost: the reported range was $1,339 for an Opus 4.8 hybrid to $10,565 for GPT-5.5 alone.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Distributed Systems](https://yomu.fyi/topic/distributed-systems), [Model Routing](https://yomu.fyi/topic/model-routing), [Rust](https://yomu.fyi/topic/rust)

- Source: [Cursor](https://cursor.com/blog/agent-swarm-model-economics)
- Source URL: https://cursor.com/blog/agent-swarm-model-economics
- Ingested by Yomu: 2026-09-02T11:42:04.263Z

[Read original post](https://cursor.com/blog/agent-swarm-model-economics)