Microsoft has rolled out a new technical architecture for its Copilot Researcher tool, moving beyond single-model inference. The system now orchestrates multiple large language models in sequence to handle multi-step research tasks. The primary new method, called Critique, first generates a response using OpenAI's GPT. That output is then passed to Anthropic's Claude for refinement and critical evaluation, creating a processing pipeline.

This design introduces a feedback loop between two distinct model families. Microsoft's engineering team suggests this approach improves factual accuracy and analytical depth by simulating a review process. Early benchmark results against systems like Perplexity's Deep Research show marginal gains in accuracy and completeness metrics.

For engineers who prioritize transparency over a single synthesized answer, Microsoft introduced a second option: Model Council. This feature runs queries through GPT and Claude in parallel, presenting both raw outputs side-by-side. It includes a diff report highlighting consensus and divergence between the models, letting the user assess reliability.

Both features are currently live in the Microsoft 365 Copilot Frontier program, serving as a testbed for production-grade multi-agent AI systems. The deployment highlights a practical trade-off: sequential processing (Critique) aims for higher quality at the cost of latency, while parallel execution (Model Council) offers interpretability and faster time-to-first-token.

Source: Engadget