# Microsoft Agent Framework Harness: What It Actually Does > Source: > Published: 2026-07-29 18:09:11+00:00 Microsoft shipped the Agent Framework Harness on July 22 — a production-ready execution layer that turns an LLM into an autonomous agent without you manually wiring up memory, planning, tool loops, approvals, and telemetry. It lands in Python and .NET. The timing is loaded: AutoGen is in maintenance mode with millions of developers hunting for a new home, and [GitHub Models goes dark tomorrow](https://github.blog/changelog/2026-07-01-github-models-is-being-fully-retired-on-july-30-2026/). Microsoft is making its move. ## What an Agent Harness Actually Is Most “agent frameworks” are elaborate wrappers. The harness is something different: it’s the execution layer between the model and the real world. A language model can only generate text. The harness is what makes it call tools, persist state, track work across multiple turns, ask for approval before taking risky actions, and stop when the job is done — rather than hallucinating its way to a false finish. Concretely, the Microsoft Agent Framework Harness ships with: **Automatic tool-calling loop**— re-invokes the model until a completion condition is met** Planning**— TodoProvider for task tracking, AgentModeProvider to separate “plan” from “execute” phases** Persistent memory**— FileMemoryProvider keeps the agent’s notes across turns in a session** Context compaction**— automatically compacts conversation history mid-loop to prevent context window overflow on long tasks** Approval flows**— human-in-the-loop with “Don’t ask again” standing rules for low-risk actions** Telemetry**— built-in observability that flows directly into Application Insights on Azure Developer input: a chat client, instructions, and optional tools. The harness configures the rest. ## The CodeAct Story Most Coverage Missed Buried in the harness release is something worth stopping on: CodeAct. When enabled, the agent collapses a multi-step plan into a single executable code block instead of executing each step as a separate tool call. Microsoft reports roughly 50% latency reduction and over 60% fewer tokens on representative workloads. The safety angle matters here too. CodeAct runs generated code inside a [Hyperlight](https://github.com/microsoft/hyperlight) micro-VM, not directly in your host environment. Model-generated code executing on your machine without isolation is a threat model most agent frameworks don’t address seriously. This one does. For teams running agents at scale, a 60% token reduction is not a footnote — it’s the difference between a viable cost structure and an expensive science project. ## What “Batteries-Included” Actually Means (and Doesn’t) Microsoft is using “batteries-included” to describe the harness, and the phrase is accurate — with limits worth knowing. **What the harness handles for you:** the tool-calling loop, planning primitives, session memory, context window management, approval rules, and telemetry wiring. This is the plumbing that every production agent needs and that every team was reinventing from scratch. **What you still own:** model selection, fine-tuning, RAG and vector store setup, and the logic inside your custom tools. The harness is runtime infrastructure, not application logic. Also worth noting: some capabilities are still in alpha. Background agents, file access tooling, and shell tooling are not yet in the stable release. Build plans around what’s GA today. ## The AutoGen Migration Context AutoGen entered maintenance mode in October 2025. The project’s README now directs new users to Microsoft Agent Framework and links an official migration guide. AutoGen had a substantial developer community — and that community is actively looking for a stable landing spot. Microsoft Agent Framework 1.0 GA shipped in April 2026, converging AutoGen and Semantic Kernel into a single supported platform. The July 22 harness release is the piece that makes it a complete, production-usable system. The 1.0 release gave you stable interfaces. The harness gives you the engine. The realistic competitive comparison in 2026 is MAF versus LangGraph, which reached 1.0 GA in October 2025. LangGraph has a wider ecosystem and more third-party integrations. MAF has a tighter .NET story, native Azure deployment, and out-of-box production defaults. If you’re on Azure or building for .NET, the calculus is fairly clear. ## Getting Started The Microsoft Agent Framework Harness is available now in both languages: ``` # Python pip install agent-framework # .NET dotnet add package Microsoft.Agents.AI ``` The [official release post](https://devblogs.microsoft.com/agent-framework/the-microsoft-agent-framework-harness-is-now-released/) includes working examples for both. The [GitHub repository](https://github.com/microsoft/agent-framework) has samples in `python/samples/02-agents/harness` and `dotnet/samples/02-agents/Harness` . The [Microsoft Learn documentation](https://learn.microsoft.com/en-us/agent-framework/agents/harness) covers provider configuration in depth. The agent framework wars in Microsoft’s ecosystem are effectively over. AutoGen is sunset, Semantic Kernel is merged in, and the harness ships the execution layer the platform was missing. Whether that’s good news depends on whether you’re bought into the Azure ecosystem — but for .NET teams building production agents, the plumbing argument just got a lot weaker.