Model packs + swarms
The right model for each step.
A swarm splits the job. A model pack is the models and tools for one kind of app.
The problem
One frontier model for every step.
Expensive. Hard to predict.
And too big a job for one agent.
What it does
A swarm for the size. A pack for the kind.
- Swarm. Break a big job into many small ones, hand them to smaller models, and let a stronger model direct the work.
- Model pack. The models and tools geared to one kind of app. We choose them for each task.
Catalog job · a different model for each step
Read scansScanned pages are re-read as images.
Vision modelTriageRules decide which pages are worth a model call.
RulesExtractModels pull the product facts.
Extraction modelsCheckA value is kept only if it is quoted on the page.
CodeReviewAgents from different model families check the draft.
Review agentsMeasured from the run's files, snapshot 2026-10-01. The review is early and paused.
Sample use
A model per job.
- Catalog. A vision model reads scans. Extraction models pull the facts. Review agents check the draft.
- Blog. A smaller local model drafts. A larger one drafts the pillar pages.
- Robotics. A swarm splits a point cloud into sections.
Questions
Quick answers
Who picks the models?
We do, task by task, as part of the quote.
Do you only use open-weight models?
No. We use what the outcome needs. In the jobs above most of the work ran on open-weight models.
Sound familiar?
Frontier bills for routine work.
The system