Raw capture — raw/agent-swarm-model-economics.md
url: https://cursor.com/blog/agent-swarm-model-economics title: "Agent swarms and the new model economics" author: Wilson Lin date_fetched: 2026-07-21 date_published: 2026-07-20
Cursor Blog, Research category. 17 min read.
The swarm has two roles organized around tree-like task decomposition. Planner agents (powered by top models) split goals into pieces and delegate. Worker agents (faster, cheaper models) execute those pieces. The swarm's shape grows to match the problem's contours rather than imposing a fixed topology.
"The design is a superset of more rigid orchestration systems."
The approach generalizes to building a browser, solving math problems, optimizing GPU kernels, finding vulnerabilities, raising test coverage, and generating synthetic training data.
A single agent handling a full task must walk the entire tree, holding ancestors, position, and goals in context simultaneously.
"We think this explains why long-running single agents drift."
In the swarm, planners never implement (context stays free of low-level detail) and workers never plan (context is fully devoted to one narrow piece). The authors cite economist Ronald Coase's theory of the firm — coordination costs grow faster than work itself, leading to bounded tiers.
The earlier browser swarm peaked at ~1,000 commits/hour on Git. The new system hits ~1,000 commits/second. Cursor built a new VCS from scratch to enable this throughput and make collisions visible.
Five failure modes identified:
Split-brain design — Two planners unknowingly implement the same concept differently. Fix: planners make decisions directly and ensure no two subtrees decide the same question.
Contention between planners — Planners aware of each other fight over the same files. Fix: shared design docs with compile-checked references; a reconciler merges contradictory docs.
Merge conflicts — Workers collide on files but can't absorb other agents' context to merge. Fix: a neutral third-party agent intervenes to resolve conflicts impartially.
Megafiles — Popular files grow bloated, becoming expensive to transport/diff/merge and sites of constant collisions. Fix: workers flag bloated files, new commits block, and an outside agent decomposes them.
Ossification — Agents avoid touching core code even when needed. Fix: "license intentional breakage" — agents can make focused patches outside scope with explanatory comments; the compiler propagates changes downstream.
"The compiler carries the change through the rest of the system."
Multiple review approaches were tested — full transcripts, output-only, codebase-only, different models/personalities.
"No single lens catches everything, but decorrelated lenses stack."
Review compute is described as high-return since it's much cheaper than the work it audits. This stacked review system is cited as a major contributor to sustained quality.
The concept of stigmergy (from ant/termite colonies) is introduced — organisms coordinate by shaping the environment, which in turn shapes the next organism. The Field Guide is a self-authored folder whose index.md is injected into every agent at start, curated by agents with only a line budget constraint.
"model weights are frozen, so it's precisely surprise encounters that are worth capturing"
The swarm was instructed to implement the entire 835-page SQLite manual in Rust. Source code, test suites, SQLite binary, and internet access were withheld. Progress was graded against sqllogictest (millions of queries with known answers). The swarm was unaware of the test suite's existence. Human reviewers checked for cheating.
Four configurations tested: 1. GPT-5.5 as both planner and worker 2. Grok 4.5 as both planner and worker 3. Opus 4.8 planner + Composer 2.5 worker 4. Fable 5 planner + Composer 2.5 worker
The new harness outperformed the old in every mix. The Fable 5 hybrid passed ~2/3 of the suite within the first hour. New runs scored 73–85% at the 4-hour cutoff; old runs scored 11–77%. Every new configuration eventually passed 100%.
GPT-5.6 Sol was originally intended but produced "runaway spirals unlike anything the other models produced" and was replaced with GPT-5.5.
The old Grok 4.5 run produced 68,000 commits in 2 hours (~70× the new run's pace), but most were thrash. The old run accumulated >70,000 merge conflicts (accelerating); the new run had <1,000 over 4 hours. The hottest file in the old run had 7,771 conflicts touched by 1,173 agents; in the new run, the most contested file saw 47.
"the old swarm's biggest coordination failure — split-brain"
The old run sprawled to 54 Rust crates (including 3 separate SQL packages); the new run settled on 9 crates. In the Fable 5 mix, the old swarm needed 64,305 lines of engine code to pass the suite; the new one used 9,908. Opus mix: 19,013 lines at 97% (old) vs 4,645 lines at 100% (new).
Costs ranged from $1,339 (Opus 4.8 hybrid) to $10,565 (GPT-5.5 alone). Workers carried at least 69% of tokens (over 90% in most runs). However, planner tokens cost more — in the Opus/Composer mix, Opus produced few tokens but ~2/3 of the cost.
"Few moments in a large task genuinely require frontier intelligence."
GPT-5.5 worker costs alone: $9,373. Opus planner + Composer 2.5 workers: entire worker fleet cost $411. Fable 5 used fewer planning tokens than Opus (despite 2× per-token price) but its workers consumed many more tokens, making the total run more expensive.
Each AI capability jump raised the abstraction level: autocomplete → line, early models → block, agents → file/feature, swarms → spec.
"What was scarce in this experiment... is the right description of intent."
The swarm is compared to a compiler — translating intent (analogous to source code) into executable work through intermediate steps. The difference: compilers preserve meaning deterministically while the swarm is probabilistic at every step. The entire post describes efforts to close that gap.
The public codebase is available at github.com/cursor/minisqlite (solo Opus 4.8 run).