Microsoft's Agent Framework Hits GA: Build & Run AI Agents

The Runtime Revolution in Agent AI

Microsoft's Agent Framework reaching General Availability is a pivotal moment for AI development, moving beyond mere libraries to a fully supported production runtime. The 'harness' concept, as highlighted, is crucial: it's the operational layer that imbues language models with the ability to interact with tools, manage multi-step tasks, and maintain state. This shift from 'building' to 'running and governing' agents is a significant step towards enterprise-grade AI deployments. The inclusion of integrated observability (OpenTelemetry), security, and policy enforcement mechanisms directly addresses the critical challenges of managing AI agents in production environments. The ability to seamlessly integrate third-party coding agents like GitHub Copilot and Claude SDKs under a unified governance model is particularly noteworthy, simplifying complex orchestration and ensuring consistency across diverse AI components.

The article effectively emphasizes that the harness is the bulk of the work, citing the MBZUAI study to underscore this point. This perspective is vital for understanding the value proposition of a standardized runtime. By providing this out-of-the-box, Microsoft is enabling platform teams to focus on agent logic and business value rather than reinventing complex infrastructure. The consumption-based billing for Foundry Hosted Agents also signals a move towards more flexible and scalable AI deployments. However, a potential concern might be the vendor lock-in associated with the 'Foundry Hosted Agents' as a managed deployment target, even though the framework itself is open-source. While the connectors abstract away some complexity, the underlying dependencies and the specifics of how policies are enforced across different agent types will be critical for long-term maintainability and interoperability. The benchmark comparison with GitHub Copilot SDK, while insightful, is also a self-comparison, and independent evaluations will be key to fully assessing performance and safety nuances across various models and scenarios.

Key Points

• Microsoft Agent Framework, including Harness and Hosted Agents, has reached General Availability (GA).
• The 'harness' is presented as the core runtime infrastructure for AI agents, enabling tool use, multi-step task execution, and state management, rather than just a library for building them.
• Key features include function invocation, history persistence, context compaction, file memory, skills, web search, tool approval, and built-in OpenTelemetry.
• Foundry Hosted Agents offer a managed deployment target with consumption-based billing.
• The framework supports seamless integration with third-party coding agents like GitHub Copilot SDK and Claude Agent SDK, bringing them under unified governance and observability policies.
• Orchestration patterns (sequential, parallel, Magentic) are stable, allowing teams to change coordination styles without rewriting agent code.
• The emphasis is on the runtime and governance layer, shifting the focus from what an agent can do to who ran it, under which policy, and where its trace lands.

---

📖 Source: Microsoft Agent Framework Harness and Hosted Agents Reach General Availability