Microsoft's latest Agent Framework push matters because it turns the production harness into framework default instead of leaving it as custom glue around the model.
What Launched
On June 3, 2026, Microsoft published its Build 2026 Agent Framework announcement, highlighting Agent Harness, Foundry Hosted Agents, CodeAct, GitHub Copilot SDK integration, and handoff-based multi-agent orchestration.
Microsoft also says Agent Framework reached 1.0 GA on April 2, 2026, positioning the stack as a supported convergence of AutoGen and Semantic Kernel across .NET and Python.
Why The Harness Layer Matters
The strongest part of the release is not branding. It is the decision to make context compaction, file memory, approval controls, skills, background agents, telemetry, and web search part of the standard harness surface.
That matters because production agents fail at the harness boundary long before they fail at prompt wording. They lose state, exceed context, call the wrong tool, or become too brittle to recover after interruption.
Why CodeAct and Hosted Agents Matter
Microsoft is also pushing the framework toward operational efficiency. Hosted Agents give teams a production home with session isolation, persistence, and observability. CodeAct reduces multi-step tool overhead by letting the model write one short program that executes inside a sandbox instead of burning many model turns on serial tool calls.
That is a practical maturity signal. The framework race is becoming a contest over runtime behavior, cost, and recovery properties, not only developer ergonomics.
The Take
Microsoft is making a clear argument that the right abstraction for serious autonomous work is not just an agent object. It is an agent plus harness plus hosted runtime. That is exactly how the zero-human company stack is starting to solidify.
Related: See our previous research on Google ADK Go 2.0, Cloudflare Flue, and AWS Step Functions AgentCore.