A runtime milestone for Microsoft’s agent stack

Microsoft has moved Agent Framework Harness and Foundry Hosted Agents into general availability, turning Agent Framework from a build-time SDK into a supported production runtime for AI agents. Agent Framework 1.0 was released on April 2, 2026. At Build 2026, held June 2–3, Agent Harness, connectors for the GitHub Copilot SDK and Claude Agent SDK, and multi-agent orchestration patterns reached stable release.

The important shift is that Microsoft is no longer positioning Agent Framework only as a library for creating agents. It is now offering a runtime layer that platform teams can use to execute, manage, observe, and govern them. Harness is delivered as a single binary that can run across local development setups, container environments, and managed deployments. Foundry Hosted Agents provide the managed target and are billed by usage.

What Harness adds around the model

Microsoft had previously framed Agent Framework as a consolidation of the open-source Semantic Kernel and AutoGen efforts. With the 1.0 release, those predecessor projects moved into maintenance mode, answering the developer question of which framework to use. The next problem is operational: where agents run, what resources they can access, and how their behavior appears in policy and observability systems.

Harness is the answer to that operational layer. In simple terms, a harness is the execution environment around a model. The model produces text and reasoning steps; the harness gives it controlled access to tools, task state, memory, approvals, telemetry, and stopping rules. Microsoft principal software engineer Wes Steyn summarized the distinction by noting that a model alone can only generate text. To invoke tools, complete multi-step tasks, and continue until a job is done, it needs to be wrapped in a runtime.

The release enables several capabilities by default:

• function calling, persistent history for each call, and context compression;
• todo lists with planning and execution modes, file memory, and skills;
• web search, tool approval, and built-in OpenTelemetry.

OpenTelemetry is an open-source observability standard used to collect traces, metrics, and logs in a consistent way. Shell tools, file access, background subagents, and automatic loops remain optional, and warnings are still shown when they are enabled. Developers supply the chat client, instructions, and tools; Harness handles planning, persistence, compression, approval, search, and telemetry through a single call.

The hard part is the system, not just the model

The emphasis on Harness reflects how much engineering sits outside the model itself. A paper published in April 2026 by MBZUAI’s VILA Lab, titled “Inside Claude Code,” examined a concrete case. Researchers analyzed the TypeScript source for Claude Code v2.1.88, which briefly became available on March 31 when Anthropic published an npm package containing source maps. They counted 1,884 files and roughly 512,000 lines of code.

Their estimate was striking: about 98.4% of the code related to harness infrastructure, permissions, context management, sandboxing, tool routing, and recovery, while AI decision logic accounted for about 1.6%. The authors cautioned that this was a line-count classification of the leaked package, including generated and minified code, not a full audit. Still, the pattern is consistent with other independently built agents, including Codex CLI and Aider. The constraint appears architectural: reliable agents need far more than a prompt and a model endpoint.

An early benchmark pointed in the same direction. Microsoft chief AI architect Aqib Sherwani compared two Microsoft runtimes, Agent Framework and the GitHub Copilot SDK, while fixing model parameters and starting with a deterministic simulation so differences could be attributed to the harness. His conclusion was that the reasoning was the same, while the engineering differed. Both reached the same answer in the same number of steps, but their runtime controls diverged. Agent Framework stopped itself after 40 round trips and returned a limit-reached message. With host-side stop control disabled, the Copilot SDK continued to 300 trips without stopping itself. In short, one runtime puts the brake inside the loop; the other expects the host to supply it.

Connectors make governance practical

The stable connectors for GitHub Copilot SDK and Claude Agent SDK let Agent Framework delegate tasks to coding agents without custom adapters. Each agent still runs its own autonomous loop, but the wrapper allows coding agents to collaborate in the same workflow with Azure OpenAI, Anthropic, or custom agents.

The governance detail matters. These connectors follow the identity, content-safety, and observability policies configured for the agent cluster. Their traffic flows into the same OpenTelemetry traces and Foundry dashboards as other agent traffic, rather than becoming a separate integration with its own access model. That is the control-plane question now facing agent platforms: not only what an agent can do, but who ran it, under which policy, and where its traces went.

Microsoft also stabilized orchestration patterns alongside the test framework. These include sequential pipelines, parallel collaboration, and the Magentic pattern derived from Microsoft Research’s Magentic-One. Its 2024 evaluation reported 38% on GAIA, 27.7% on AssistantBench, and 32.8% on WebArena. Microsoft said the first two results were statistically comparable with the state of the art at the time, while WebArena was competitive. Because the patterns share a common API, teams can change coordination styles without rewriting agent code.

The agent market is moving toward runtime competition

For platform teams, the cloud significance of this release is the runtime rather than the SDK. Microsoft is offering a supported governance framework, a usage-billed hosted target, and a policy and observability model that treats third-party coding agents as managed members of a cluster.

The framework, governance layer, and connectors are now available on GitHub for .NET and Python, lowering the barrier to experimentation. The larger direction is clear: production agents are starting to resemble cloud-native applications. The model is only one component. What determines whether an agent can be deployed safely is the surrounding runtime: permissions, audit trails, stop controls, telemetry, and orchestration. With Harness and Hosted Agents, Microsoft is trying to standardize that layer before it becomes the main battleground for enterprise agent platforms.