Deploy. Observe. Learn. Reinforcement learning for production agents | BRK231

Author: Microsoft Developer - Bewertung: 0x - Views:9

Agents don't fail in demos — they fail in production. In this breakout, see how teams use fine-tuning and reinforcement learning on Microsoft Foundry to improve production agents using real usage signals. We cover when fine-tuning reduces cost and latency, when RL delivers deeper gains, and how Foundry makes it easy to train, evaluate, and redeploy safely.



Seating for this session is first-come, first-served. Add it to your schedule to plan your day and arrive early to secure a spot.



To learn more, please check out these resources:

* https://aka.ms/build26/BRK231

* https://aka.ms/build/foundrydiscord





𝗦𝗽𝗲𝗮𝗸𝗲𝗿𝘀:

* Alicia Frame

* Omkar More





𝗦𝗲𝘀𝘀𝗶𝗼𝗻 𝗜𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻:

This is one of many sessions from the Microsoft Build 2026 event. View even more sessions on-demand and learn about Microsoft Build at https://build.microsoft.com



BRK231 | English (US) | Working with models





Breakout | (200) Intermediate





#MSBuild



Chapters:

0:00 - Technical Breakdown: How Fine-Tuning Works

00:09:00 - Transition to Demo: Cheaper and Faster Models with Distillation

00:10:10 - Demo Setup: Retail Customer Service Agent Scenario

00:13:44 - Evaluating with traces and experimenting with smaller models

00:14:34 - Setting up Foundry graders for evaluation

00:26:16 - Configuring RFT grader and avoiding reward hacking

00:33:02 - Introducing the new interactive training API for custom fine-tuning control

00:36:36 - Introducing tool invocation and control in sampling process

00:37:05 - Explaining GRPO step and gradient computation