Skip to content

BRK264 Breakout Wed 18 Nov 15:00 PST

Pick the right harness to build your agents with using Copilot Studio

Alice Appleton, Jason Xian, Jennifer Ratten

Summary

From the recording / Microsoft AI summary

The session described three harnesses: Microsoft Copilot (complex, long-running processes, always consumption-billed), customer conversation (voice and chat, low latency) and standard (kept for continuity). Microsoft's offline evaluations claim input tokens down 45% and pass rate up 10 points with OpenAI models and 22 with Anthropic models. A customer example kept humans in the loop for exceptions. A slide estimated that premium reasoning models complete 2.4x more tasks and use 87% fewer credits on the Microsoft Copilot harness, based on Microsoft's offline evaluation dataset.

Session description

Microsoft's catalogue description

When should you not build using the Microsoft Copilot harness in Copilot Studio? Come learn about how factors like the business process, its complexity, user experience, and cost influence which harness you would want to build your agent with. Leave the session knowing how to match the right harness to the job, while building agents that are fit for their intended purpose.

Demo timestamps

Microsoft's player does not support links to a specific time, so open the recording and skip to the time shown.

  1. Skip to 00:18:46

    Live agent run: draft client email with recommendations

    After a comparison of the standard harness and the Microsoft Copilot harness, a draft email the agent prepared for a client is opened. It includes recommendations where the portfolio is not meeting agreed goals and guardrails.

  2. Skip to 00:20:22

    Agent run completes for under a dollar

    The speaker returns to the agent and notes the run completed for less than a dollar, then hands over to a customer story.

  3. Skip to 00:37:37

    Customer-conversation harness demo shown as recording

    A live run is not possible because of poor reception backstage, so a recent demo recording of the customer-conversation harness is played instead.

Watch on-demand (opens in a new tab)

← All sessions