Semantic Kernel Integrated AI Into Enterprise Code. The Agents It Creates Have No Schema.

Semantic Kernel's approach to enterprise AI integration is well-considered. Treating LLM capabilities as native functions, providing strongly-typed plugin interfaces, and supporting multiple AI services through a unified API are genuine contributions. The gap described here is not about the SDK. It is about the agents the SDK produces.

Plugins are capabilities, not agent fields

Semantic Kernel agents are assembled from plugins (functions the agent can call), memory (conversation and semantic memory connectors), and planners (strategies for composing function calls). The kernel orchestrates these components to respond to user requests.

But plugins are capability descriptors, not agent schema fields. A Semantic Kernel agent has no typed governance field. It has no policy reference that constrains its behavior. It has no integrity metric that tracks behavioral consistency. The agent's structure is whatever the developer composes. There is no canonical schema that says: this is what an agent must be.

Memory connectors are not semantic memory

Semantic Kernel provides memory connectors for conversation history, semantic search, and vector stores. These are storage abstractions, not semantic memory with governance.

In a canonical agent schema, memory is a typed field with defined structure, lineage tracking, and governance constraints. Every memory entry has provenance. Memory mutations are validated. Memory persists across execution cycles with full lineage. Semantic Kernel's memory connectors provide storage. They do not provide governed semantic memory.

What a canonical agent schema provides

A canonical agent schema defines typed fields that constitute an agent: identity, memory, governance, capabilities, execution state, and lineage. Semantic Kernel's plugins would map to the capabilities field. Its memory connectors would implement the memory field. But the governance, identity, and lineage fields would be new: structural components that make the agent validatable and governable.

With a canonical schema, Semantic Kernel agents deployed in enterprise environments would be structurally interoperable. An agent built by one team could be validated by another team's governance framework because both share the same schema. Compliance becomes structural rather than procedural.

The remaining gap

Semantic Kernel made AI integration natural for enterprise developers. The remaining gap is in agent definition: a canonical schema that makes agents structurally defined, governable, and interoperable objects rather than ad hoc compositions of plugins and memory connectors.