Claw and agent harness in Microsoft Foundry | BRK243
Shawn Henry, Amanda Foster, and Glenn Condron go deep on building and operating multi-agent systems on Microsoft Foundry, focusing on “agent harness” patterns (including Claw) and hosted agents architecture. They cover long-running agents with triggers, state and file access, plus how Copilot SDK and Claude Agent SDK fit into coordinated workflows.
Overview
This Microsoft Build 2026 breakout (BRK243) focuses on designing and running multi-agent systems on Microsoft Foundry (Azure AI Foundry). It highlights:
- Claw agent patterns and an agent harness approach for coordinating agents
- Hosted agents architecture on Foundry
- Support for long-running agents with:
- Triggers
- State management
- File access
- How coding agents built with:
- GitHub Copilot SDK
- Claude Agent SDK integrate into multi-agent workflows using Microsoft Agent Framework
- Operational concerns for these systems, including:
- Observability
- Telemetry
- Continuous evaluations (evals)
Session chapters (from the video description)
- 0:00 - Definition and purpose of agent harnesses
- 00:05:51 - Key components of agent harnesses: context, tools, memory, lifecycle
- 00:13:18 - Explanation of Hermes file generation and context management
- 00:16:08 - Automated nightly maintenance, backups, and routine creation in Hermes
- 00:20:11 - Practical setup using Foundry components for Hermes customization
- 00:27:21 - Customizing research agent setup with tools, memory, and telemetry
- 00:28:18 - Demonstrating harness console interaction and human-in-the-loop planning
- 00:31:31 - Transition to GUI-based agent operation using AGUI and Copilot Kit
- 00:41:06 - Autopilot workstream manager agent and group chat behaviors