Building AI Agents with Microsoft Foundry Agent Service
Microsoft Foundry Agent Service introduces the Copilot harness, a managed execution system for building AI agents that reason, use tools, and complete complex, long-running tasks. It abstracts away common orchestration machinery, allowing developers to focus on an agent's expertise and evaluation. The service also offers general availability of long-running resilience for hosted agents, providing recovery foundations for workflows that may span hours or days.
Read original on Azure Architecture BlogManaged Execution for AI Agents
Building robust AI agents often requires implementing complex orchestration logic for coordinating model and tool calls, managing context, and handling multi-step tasks. Microsoft's Copilot harness in Foundry Agent Service provides a managed execution system that abstracts these complexities. This allows developers to define agents declaratively, focusing on the agent's core logic, skills, and tools rather than the underlying infrastructure. The service manages the agent loop, sandboxing, and coordination of steps toward a desired outcome.
Key Building Blocks of the Copilot Harness
- Reusable Skills: Package instructions, scripts, and resources for agents to apply across different tasks.
- Connected Tools: Integrate supported Foundry and Model Context Protocol (MCP) tools with configured authentication and approvals.
- Execution and Delegation: Agents can interact with files, execute code, and delegate specific tasks to subagents while a main agent coordinates results.
- Source-controlled Agent Definitions: Manage agent definitions and resources alongside application code, supporting creation and testing within Foundry or external tooling.
The harness is designed to manage working context dynamically as tasks evolve and to coordinate model and tool interactions within configured limits and policies. This approach enables teams to focus on improving agent expertise and evaluating results, leading to faster development cycles.
Long-Running Resilience for Hosted Agents
For workflows that require custom frameworks or execution logic, Foundry Agent Service offers hosted agents. A significant enhancement is the general availability of long-running resilience, which allows hosted agents to recover from interruptions during tasks that can last hours or days. Foundry detects interruptions and restarts the work with the original input. Developers are responsible for saving progress and intermediate results so the agent can resume from the last checkpoint. This also implies careful consideration of idempotency for external system updates to prevent duplicate actions upon restart.