Letting Models Choose Subagents? The Hidden Pitfalls in Agent Orchestration

davidcrawshaw · x · 2026-07-31

A developer building multi-agent systems highlights a tricky limitation: letting the primary model autonomously choose which subagent model to use yields poor results. For instance, GPT-4o might blindly defer to Claude Sonnet for critical architectural decisions. This indicates that routing in complex agent workflows still requires explicit human guidance.

Original post →

More from coding & agent

coding & agent channel →