# Cursor Agent Swarms
In July 2026 [[Cursor.com|Cursor]] published the numbers behind its agent swarm work, and it is the most useful public document on [[AI Agent Swarms|agent swarm]] economics so far. Not because the idea is new, but because they ran the same task through several model combinations and published what each one cost.
The task: implement SQLite from scratch, measured against the sqllogictest suite.
## The economics
Planner agents run on the smartest models and split a goal into pieces. Worker agents run on faster, cheaper models and execute those pieces. The argument for the split is one sentence: **few moments in a large task genuinely need frontier intelligence.** The decomposition, the design decisions, a handful of trade-offs. Everything after that is following instructions, and following instructions is cheap.
The cost table makes the point better than the argument does:
| Configuration | Total cost |
|---|---|
| [[Claude Opus 4.8\|Opus 4.8]] planner + Composer 2.5 workers | **$1,339** |
| Grok 4.5 for both roles | ~$2,500 to $3,000 |
| [[GPT-5.5]] for both roles | **$10,565** |
The line worth staring at: GPT-5.5 workers alone cost $9,373. Opus planning plus Composer workers cost $411 for the same worker layer. **The worker tier is where the money goes**, because that is where the volume is, and it is the tier where model choice matters least for quality.
This is the same recipe as the cost-aware orchestration pattern in [[Claude Fable 5]], with a price tag attached.
## What actually broke, and what fixed it
The interesting half of the post is not the economics, it is the failure catalogue. Same model, same task, old orchestration versus new:
| | Old swarm | New swarm |
|---|---|---|
| Commits | 68,000 | 1,000 |
| Merge conflicts | 70,000+ | under 1,000 |
| Crates | 54 (3 duplicate SQL packages) | 9 stable |
| Worst-contested file | 7,771 conflicts | 47 |
| Final engine code | 64,305 lines | 9,908 lines |
Six times less code for the same working result. The old swarm was not slower, it was **generating waste at industrial scale** and calling it progress.
The named failure modes and their fixes:
- **Split-brain**: two agents build the same thing. Fixed by making planners write decisions into shared design docs with compile-checked references. Coordination through artefacts, not through conversation
- **Merge conflicts**: resolved by a third-party reconciler agent, deliberately impartial, belonging to neither side
- **Megafiles**: workers flag files that have bloated, and an external agent decomposes them
- **Ossification**: agents become afraid to touch core code. Fixed by explicitly permitting justified patches with explanatory comments
- **Review**: multiple decorrelated lenses stacked. Full transcript, output only, codebase only. Same principle as never letting a worker check its own work
- **The Field Guide**: a shared context file the agents write themselves, curated by them, capturing surprising learnings and injected into every agent
They also built a custom VCS, because Git handles about 1,000 commits per hour and they needed 1,000 per second.
## Why it matters
**Context efficiency beats raw parallelism.** This is the sentence to keep. The old swarm had more agents doing more work and produced a worse result more expensively. Separating planning from execution is what made scale usable, not adding workers.
**Their framing is that a swarm resembles a compiler**, lowering high-level intent through intermediate stages into executable work. That is a better mental model than "many agents working together", because it tells you the stages should be typed and checked rather than chatty. It rhymes with the case for [[DSLs Make LLM Output Reliable|DSLs as handoff contracts]] and with the node-contract rule in [[Graph Engineering]].
**Every fix here is topological.** Impartial reconciler, decorrelated reviewers, shared design docs, self-curated field guide. None of them make any individual agent smarter. That is exactly the claim graph engineering makes, now with a control group.
## Caveats
- Cursor benchmarked Cursor, and Composer is Cursor's own model. The cheap-worker conclusion happens to favour their product. It is also the same conclusion practitioners reached independently, which is the reason to take it seriously
- The full model matrix was not tested, only selected combinations
- Manual review found no test-specific shortcuts, but they say plainly they have not done a deeper analysis of the output
- Agents chose their own strategies, so timings vary between runs
- One task, one domain. SQLite has an unusually good oracle (a giant conformance suite). Most real work does not, and the whole approach leans on having one
## References
- Cursor, "Agent swarms and model economics" — https://cursor.com/blog/agent-swarm-model-economics
## Related
- [[Cursor.com]]
- [[AI Agent Swarms]]
- [[Challenges in Managing AI Agent Swarms]]
- [[AI Agent Orchestration]]
- [[Graph Engineering]]
- [[Loop Engineering]]
- [[DSLs Make LLM Output Reliable]]
- [[Claude Fable 5]]
- [[Claude Opus 4.8]]
- [[GPT-5.5]]
- [[AI Agents]]
- [[Agentic Engineering]]
- [[How Coding Agents Work]]
- [[AI Verifiability]]
- [[Kimi K3]]
- [[Software Architecture]]