Cohere AI Updates: October 15, 2026

1. Cohere Research Finds Delegating Planning Is the Cheapest Way to Scale Coding Agent Teams

Cohere. Cohere researchers tested lead-and-worker coding agent teams in a minimal bash-only harness modeled on mini-SWE-agent, where a lead splits work through a delegate_agent tool, on the 200-task ProgramBench benchmark (rebuilding programs by probing an executable) and on Commit0-Lite. With GPT-5.6 Sol, teams beat solo agents (63.1% vs. 57.7% with an explore/plan, implement, verify prompt) but cost 4 to 7 times as much, and delegating only the planning stage reached 63.6% at $249 compared with 63.2% at $510 for delegating all three stages. A Sol lead paired with 16 cheaper GPT-5.6 Luna planners scored 61.8% at $155, about 70% below the all-Sol team, but cheap workers could not compensate for a weak lead: Luna, Terra, and Sol leads with Luna workers scored 42.5%, 51.6%, and 60.5%. Worker-to-worker messaging did not help stronger models and cost Luna teams 4 to 7 points, and on a 30-task subset gains varied widely by model, from 6.9 points for GPT-5.6 and 5.9 for Claude Opus 4.8 to none for Haiku 4.5 and Gemini 3.7 Flash. For practitioners building multi-agent coding systems, the results suggest putting the strongest model in the lead role and spending extra workers on exploration rather than implementation. Source