Designing an agent harness that orchestrates models | BRK280

Aashna Garg, Andy Salerno, and Anisa Majhi present an advanced session on designing agent harnesses that decide how and when to use different models. The discussion focuses on routing decisions and the trade-offs involved in building adaptable agent systems.

Model routing and orchestration

The session shifts the focus from choosing a single model to coordinating models across tasks, constraints, and stages of development. An agent harness provides the orchestration layer for those decisions.

The presenters examine the signals that guide routing and how to balance four considerations:

Coordinating models as a portfolio

GitHub Copilot Fusion is discussed as an approach to coordinating models as a portfolio rather than relying on one model choice. The session offers a framework for thinking about adaptable agent systems and the trade-offs that shape their routing decisions.

Session details and resources