Microsoft has released Agent Framework as a supported production runtime, moving past the build-stage SDK into infrastructure that can run and govern AI agents in live environments. The framework hit 1.0 general availability on April 2, 2026, according to an InfoQ report published in August 2026. Build 2026, which took place June 2 to 3, brought the Agent Harness, GitHub Copilot SDK and Claude Agent SDK connectors, and multi-agent orchestration patterns to stable release, with the harness and Foundry Hosted Agents now generally available for teams to deploy.
The runtime ships with function invocation, per-call history persistence, context compaction, a to-do list with plan and execute modes, file memory, skills, web search, tool approval, and built-in OpenTelemetry, each turned on by default and individually removable. Shell tooling, file access, background sub-agents, and automatic looping stay opt-in and still trigger a warning when switched on. Foundry Hosted Agents, the managed deployment option, charges based on consumption. The harness runs as one binary across local development, containers, and hosted deployment, and the 1.0 release settled the build-time question of which framework to choose, placing both predecessors—Semantic Kernel and AutoGen—into maintenance mode.
Research from MBZUAI's VILA-Lab, published in April 2026, found that roughly 98.4% of Claude Code v2.1.88's codebase is harness infrastructure—permissions, context management, sandboxing, tool routing, and recovery—while about 1.6% is AI decision logic. The researchers examined the full TypeScript source, briefly visible on March 31 when Anthropic shipped an npm release with a sourcemap bundle, classifying around 512,000 lines across 1,884 files. Microsoft principal software engineer Wes Steyn wrote that "a model on its own can only generate text," and wrapping it in a runtime lets it call tools, work through multi-step tasks, and keep going until a job finishes. An early benchmark from Microsoft AI principal architect Aqib Sherwani, holding the model fixed and running a deterministic mock first, showed Agent Framework halted its own loop after 40 round-trips and returned a limit-reached message, while the Copilot SDK ran to 300 without stopping on its own when host-side stopping controls were off.
The Agent Framework orchestration can hand off to the GitHub Copilot SDK or the Claude Agent SDK without custom adapters, each running its own autonomous loop but wrapped so a coding agent fits alongside Azure OpenAI, Anthropic, or custom agents in one workflow. The connectors respect the identity, content safety, and observability policies already set for the fleet, according to the report, putting coding-agent traffic into the same OpenTelemetry traces and Foundry dashboards as everything else instead of creating a separate integration with its own access model. The orchestration patterns—sequential pipelines, parallel collaboration, and the Magentic pattern from Microsoft Research's Magentic-One—share one API, so teams can change coordination styles without rewriting agent code. Magentic-One's 2024 evaluation reported 38% on GAIA, 27.7% on AssistantBench, and 32.8% on WebArena, statistically comparable to the state of the art on the first two and competitive on WebArena.
For platform teams weighing the release, the Cloud-relevant takeaway is the runtime, not the SDK: a supported harness, a consumption-billed hosted target, and a policy and observability model that handles third-party coding agents as governed members of the fleet, the report says. The framework, harness, and connectors are available now in .NET and Python on GitHub.