Kye Gomez (swarms) · 2026-10-19 20:55 UTC

The best model for a task isn't always the best teammate. A new benchmark just dropped testing which models are best at collaborating with other AI agents across different types of tasks. The findings are fascinating. GPT-6 Astra leads overall at 78.0, followed by GPT-6.1 Sol at 74.1 and Claude Opus 5.5 at 70.4. But the biggest takeaway is that collaboration is still a hard problem, even for frontier models. Models are strong at resisting unsafe requests and reporting misconduct, but struggle with conflict resolution, emotional intelligence, taking initiative, and maintaining the integrity of a session when a router switches models mid-task. That last finding is especially important for anyone building multi-agent systems. A model switch can inherit more than context. It can inherit a previous model's mistakes, manipulated tests, or false records. I just built a skill for teaching model routers how to make smarter routing decisions based on these findings. Instead of routing only on cost, latency, or individual model capability, routers should account for the kind of collaboration a task requires and the risks associated with handing work from one model to another. The core lesson is the best model for a task is not always the best teammate. Skill link is below ⬇️

Loading the story on Tree News…