Microsoft Stabilizes Agent Harness, Claude SDK Included

Microsoft stabilized the Agent Harness in its Build 2026 Agent Framework release, per the framework's devblog.

The harness covers shell and filesystem access, human-in-the-loop approval, and context handling across long sessions, now at the same stability tier as skills and memory. For teams building Claude-based agents on Microsoft's infrastructure, that's the difference between an experimental API and one Microsoft has committed to keeping stable. The Claude Agent SDK integration now sits alongside Microsoft's own GitHub Copilot SDK integration in the framework.

Foundry Agent Service adds Hosted Agents on top of that. It packages your agent code as a container on managed infrastructure, handling identity, autoscaling, session state, observability, and versioning, per the devblog.

MAF hit 1.0 GA on April 2, converging AutoGen and Semantic Kernel into a single framework instead of two competing ones, per the devblog. The harness work extends that consolidation from the framework layer into how agents actually run once deployed.

Microsoft would rather host Claude-built agents on its own infrastructure than lose that workload to Anthropic. That bet holds only as long as Foundry stays the easiest place to deploy them. Watch whether the Claude Agent SDK integration stays this deep once Hosted Agents becomes the default way people ship agents on Foundry.