Agent Framework v1.0

Microsoft released Agent Framework 1.0 on April 3, shipping a production-ready toolkit for building and running artificial intelligence agents in.NET and Python. (devblogs.microsoft.com) An artificial intelligence agent is software that can take a goal, call tools, keep state across turns, and sometimes hand work to other agents instead of answering in one shot. Microsoft’s framework packages those pieces into one open-source stack with stable application programming interfaces, long-term support, and support for multiple model providers. (learn.microsoft.com, devblogs.microsoft.com) The 1.0 release says agents can connect to tools through Model Context Protocol, or MCP, and interoperate across runtimes through agent-to-agent, or A2A, links. Microsoft also says workflows can route work across multiple steps with checkpointing and human approval points. (devblogs.microsoft.com, learn.microsoft.com) Microsoft positions the project as a merger of two earlier lines of work: Semantic Kernel, its enterprise software development kit, and AutoGen, its research-driven multi-agent system. The company introduced the framework in October 2025 and said the 1.0 release followed a release candidate in February 2026. (devblogs.microsoft.com, devblogs.microsoft.com) That timeline matters because agent tooling has moved from demos toward deployment, where companies need logs, approvals, and predictable interfaces as much as model quality. Microsoft’s overview page says the framework includes middleware, session state, memory hooks, and telemetry alongside the model calls. (learn.microsoft.com, devblogs.microsoft.com) The browser-based DevUI is part of that push, but Microsoft describes it as a development tool rather than a production console. Its documentation says the app provides a web interface for interactive testing, an OpenAI-compatible application programming interface backend, and OpenTelemetry traces for debugging agent behavior. (learn.microsoft.com) Microsoft’s own docs draw a line between an agent and a workflow. Use an agent for open-ended, conversational work with tool use; use a workflow when the steps, routing, and execution order need explicit control. (learn.microsoft.com) The framework also spreads beyond Microsoft’s own model stack. Microsoft says Agent Framework supports Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI, Ollama, and other providers, which lets developers swap models without rewriting the orchestration layer. (learn.microsoft.com) Security is still a moving target inside the broader project. The main 1.0 announcement centers on stable APIs and interoperability, while the public GitHub repository shows newer Hyperlight-based sandboxing work in alpha rather than as a fully mature core feature. (devblogs.microsoft.com, github.com) By late April, Microsoft and product staff were framing observability and monitoring as part of the baseline for deployed agents, not an extra. The 1.0 release makes that argument in code: one stack for building agents, wiring them to tools, and watching what they do before those systems reach production. (devblogs.microsoft.com, learn.microsoft.com)