Multi-model orchestration: splitting one agent turn across several LLMs
Instead of one model running a ReAct loop, this design splits a turn into stages - a planner that only calls tools, small extractors that fill arguments, and a composer that alone writes user-facing text, with deterministic checks between stages. The key condition is that handoffs are written down as files, not passed as chat summaries. Multiple models only pay off when each stage has a genuinely different job.