Building agents with Claude in Foundry: Choosing your stack | THRSP1157
Maya Nielan explains how to choose a stack for production-grade Claude agents, drawing on lessons from building Claude Code and Cowork. The session focuses on which parts of the agent to own and which to delegate to a managed platform.
Separating the architecture decisions
The session organizes agent architecture into three choices:
- Model: The model capabilities that power the agent.
- Harness: The software layer that coordinates the agent’s workflow and interactions with tools.
- Infrastructure: The environment in which the agent runs and is hosted.
Separating these decisions provides a way to compare platforms without treating the model, orchestration layer, and hosting environment as a single choice. Claude’s agent harness can be used with Microsoft Foundry; the broader decision is how much control the application needs versus how much hosting responsibility a managed service should take on.
Agent stacks and deployment options
Nielan covers Anthropic’s Messages API, Claude Agent SDK, and Claude Managed Agents, alongside Microsoft Agent Framework, Foundry Agent Service, and Copilot Studio.
The comparison centers on control versus managed hosting. It considers harness and deployment options together, with examples spanning education, insurance, engineering, and enterprise use cases.
The session also addresses the transition from prototypes to production, including:
- Evolving model capabilities and what constitutes an AI agent.
- Tools, sandboxes, memory, and autonomous workflows.
- A “brain-and-hands” architecture for resilient, long-running agents.
- The role of prompts, skills, MCP, and integrations with existing systems.
Keeping tools portable
MCP and skills are presented as ways to keep tools portable across agent stacks. This portability is part of the stack-selection discussion: prompts, tools, and system integrations matter alongside the choice of model, harness, and hosting platform.
Session chapters
- 00:00 — Introduction to building production-grade AI agents with Claude and Foundry
- 00:36 — Challenges of moving AI agents from prototype to production
- 01:25 — Defining AI agents and how model capabilities have evolved
- 03:12 — Advances in tools, sandboxes, memory, and autonomous agent workflows
- 05:11 — Framework for agent architecture: model, harness, and infrastructure
- 07:27 — Anthropic developer products: Messages API, Agent SDK, and Managed Agents
- 10:39 — Brain-and-hands architecture for resilient long-running agents
- 13:39 — Prompts, skills, MCP, and system integrations
- 14:26 — Comparing Anthropic and Microsoft agent platforms and deployment options
- 19:02 — Deployment strategies for education, insurance, engineering, and enterprise use cases