Regarding Semantic Kernel Deprecation

[geffzhang]

1. Fundamental Transformation of Microsoft's AI Ecosystem: Evolution from Semantic Kernel to Agent Framework
   With the rapid evolution of AI agent technology, the mainstream development frameworks within the .NET ecosystem are undergoing a profound paradigm shift. Following Microsoft's official release of Semantic Kernel (SK) version 1.77.0, Microsoft formally announced a comprehensive migration from Semantic Kernel to the Microsoft Agent Framework (MAF). This transformation not only marks a rebranding of SK but also represents a significant leap in its underlying design philosophy—shifting from a "kernel and plugin-centric" model to an "agent and workflow-centric" architecture.

According to Microsoft's Pull Request #13852 merged in version 1.77.0 (.NET: Semantic Kernel -> Agent Framework - Migration/Samples Update), the underlying migration and sample updates are fully compatible with the new MAF 1.0 specifications. Microsoft Agent Framework 1.0, as the modern enterprise-grade successor to Semantic Kernel, provides more stable and long-term supported APIs. MAF deeply integrates multi-agent orchestration (such as sequential, concurrent, group chat, and collaborative handoff modes) with graph-based deterministic workflows. This evolution fundamentally transforms how .NET developers build autonomous agent systems, enabling them to move away from Semantic Kernel's relatively cumbersome pipeline and filter structures in favor of more intuitive, standard-aligned high-level abstractions.

Against this backdrop of technological iteration, for the openclaw.net project—an agent gateway that is independent and highly compatible with NativeAOT compilation technology—maintaining lightweight core code and high security has become the primary task in architectural evolution. This official shift directly validates the core discussed in openclaw.net's Issue #13: whether it is necessary to continue maintaining large and complex Semantic Kernel adapter code within the project, given that Microsoft has fully shifted its focus to MAF.

To maintain backward compatibility with the legacy SK in openclaw.net, the project had to maintain complex adapter code in its early stages and even release a dedicated interoperability host package, OpenClaw.Net.Runtime.OpenClaw.SemanticKernelInteropHost. This approach significantly compromised the purity and trimming safety of the runtime code. Since openclaw.net provides first-class, optional Microsoft Agent Framework adapters that allow direct switching to MAF through simple configuration (e.g., Runtime.Orchestrator=maf), completely abandoning and removing the outdated SK adapter is not only feasible but inevitable now that Microsoft has transitioned to MAF 1.0.

2. Technical Feasibility and Necessity Assessment of Deprecating the Semantic Kernel Adapter

The assessment indicates that, following the official release of Microsoft Agent Framework 1.0 and the comprehensive migration of SK version 1.77.0 (via PR #13852) to MAF, the recommendation to completely remove the Semantic Kernel adapter code from openclaw.net has a high degree of technical validity.

First, regarding Runtime Selectors, openclaw.net inherently supports Runtime.Mode (configurable as aot, jit, or auto) and Runtime.Orchestrator. If the focus shifts entirely to MAF implementation, the selection path for Runtime.Orchestrator can be streamlined from a complex multi-compatibility mode to a clearer dual-track system: native (the project's self-developed high-performance native orchestrator) and maf (Microsoft Agent Framework orchestrator). This significantly reduces the logical branching complexity during system initialization.

Second, in terms of multi-agent collaboration and long-running task orchestration, the project already supports highly reliable persistent workflow delegation through standard workflow backends (e.g., maf-durable-http). Solutions that previously relied on the legacy SK Process framework or specific pipelines for multi-agent state synchronization no longer hold a competitive advantage against MAF's first-party workflow orchestration and atomic Blackboard patterns. Retaining the old adapter would instead introduce additional design fragmentation, causing format loss when state is transferred between different adaptation layers.

Furthermore, another core subproject within the clawdotnet organization—SharpClawCode (a .NET 10 coding agent harness)—has already fully adopted the Microsoft Agent Framework in its underlying architecture, deeply integrating tool execution, permission control, session persistence, and MCP (Model Context Protocol) plugins. This consistency in technology stack means that if openclaw.net removes the legacy SK adapter and fully aligns with MAF, it can greatly facilitate the seamless reuse of common components across the organization (such as the A2A protocol and custom toolchains), accelerating the consolidation of the technology ecosystem.

Technical analysis shows that Microsoft's decision to merge PR #13852 in version 1.77.0 and formally migrate to MAF has completely eliminated the technical value of continuing to maintain backward compatibility with the legacy Semantic Kernel in new AI agent frameworks. For openclaw.net, the vision proposed in Issue #13—"focus on implementing with MAF without the need to consider legacy Semantic Kernel adapters"—is fully validated.