Prime Intellect trains Ramp spreadsheet subagent

This is the right playbook for agent products: stop waiting for a better frontier model and build a workflow-specific training loop around the bottleneck. Spreadsheet retrieval is a strong wedge because it is repetitive, measurable, and expensive enough to justify specialization.

• –Turning real user stories into synthetic tasks gives Ramp a scalable way to generate training data without relying on organic traffic alone
• –The claimed gains over Opus 4.6 show how a narrower subagent can win on latency, cost, and accuracy at the same time
• –Prime Intellect is selling the full loop here: environment design, hosted training, evaluation, and deployment in one stack
• –The broader implication is that enterprise AI may increasingly look like many small specialist agents instead of one general model doing everything