Cohere paper: 17-agent budget team matches 5 strong agents at 70% lower cost
rohanpaul_ai · x · 2026-10-19
- A Cohere research team published an arXiv paper formulating multi-agent team design as a compute allocation problem: where in the workflow should a lead delegate, which models should fill the team, and how many workers reach the reward-cost Pareto frontier.
- Key numbers: on ProgramBench, agent teams beat a solo agent but cost 3-7x more, with duplicate work and merge overhead as team size grows. With the right allocation, a heterogeneous 17-agent team comes within 1.4 points of a strong homogeneous 5-agent team at 70% lower cost.
- Three findings: planning is the most valuable stage to delegate—a lead that only delegates planning and implements/verifies itself matches a fully delegating lead at half the cost; cheap models pay off as planners; and blindly adding agents hits clear diminishing returns.
More from coding & agent
- Cua launches Spaces SDK to run Claude Code and any agent on compute you choose — Scobleizer · 2026-10-20
- Dev: top AI agents do more backup and verification on DB migrations than 99.99% of engineers — tristanbob · 2026-10-20
- 20 questions to answer before letting agents touch production data — Then_Respect_1964 · 2026-10-20
- Baseten webinar: how to evaluate open models for coding agents on your own codebase — baseten · 2026-10-20
- Every's Slack agent turns article links into videos with a storyboard-first workflow — every · 2026-10-20
- Don't be a meat proxy for your agent — make it a metal proxy for you — danshipper · 2026-10-20