Cohere paper: cheap models as planners cut multi-agent coding costs 70% with near-equal reward
rohanpaul_ai · x · 2026-10-19
A new Cohere Labs paper, Scaling Inference-Time Compute with Teams of Coding Agents, frames multi-agent team design as a compute allocation problem: a lead agent delegates subtasks to workers while varying worker role (planning/implementation/verification), model, and count.
Key findings:
- Agent teams beat solo agents on ProgramBench but cost 3-7x more, with duplicate work and extra code to merge.
- Workers used only for planning matched full-stage workers at half the cost; a strong GPT-5.6 Sol lead with 16 cheap Luna planners came within 1.4 points of an all-Sol team at 70% lower cost.
- Cheap models pay off as planners (recovering 47.5% of the reward gain for 13.9% of the extra cost) but hurt as implementers. Planner returns keep rising with team size, while implementers plateau beyond four (code fragmentation) and verifiers add little new information.
Takeaway: put your strongest model in the lead and use cheap models only for planning.
More from coding & agent
- Editor says Grok handled 90% of a music video edit — shares her 3 workflow tips — chaitu · 2026-10-20
- OpenAI Devs teases new Codex showcase: cloud environments and terminal chat — OpenAIDevs · 2026-10-20
- 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
- Build a Weekly Report Agent: Day 1 of an Agent-a-Day Series for Scientists — nlarusstone · 2026-10-20
- 20 questions to answer before letting agents touch production data — Then_Respect_1964 · 2026-10-20