BRK280 Breakout Wed 18 Nov 13:45 PST
Designing an agent harness that orchestrates models
Aashna Garg, Anisa Majhi, Andy Salerno
- Microsoft session page (opens in a new tab)
- Watch on-demand (opens in a new tab)
- Slides: rights-protected, available via the Microsoft session page (opens in a new tab)
Summary
From the recording
The session covers GitHub Copilot Fusion, which routes prompts across models using single, cascade and critique workflows. It launched as public preview in Copilot CLI this summer and became generally available on 28 October. It offers standard, high and max effort levels and works across six clients. Microsoft says there is no extra Fusion charge, and models routed through it carry a 30% discount until 30 November. Admins can apply restricted model policies. Microsoft reports Fusion Max at 71.2% on Terminal Bench 4.1. Caveat: benchmark and cost figures are Microsoft-reported; no pricing beyond the discount period was stated.
Session description
Microsoft's catalogue description
Choosing a model is becoming less important than deciding how and when models are used. Explore how routing can match models to different tasks, constraints, and stages of development. We’ll examine the signals and tradeoffs that guide these decisions, and approaches such as GitHub Copilot Fusion that coordinate models as a portfolio. Leave with a framework for building adaptable agent systems across quality, latency, cost, and control.