AIエージェント基盤のアーキテクチャを10層で整理する：OSS/SDK比較のための地図 AIエージェント基盤のアーキテクチャを10層で整理する：OSS/SDK比較のための地図

As AI agent frameworks and SDKs continue to proliferate, teams are finding it increasingly difficult to compare and evaluate their options on equal footing. A tool might get praised as "promising" in a review meeting, yet if nobody agrees on what layer of the stack it actually addresses, the discussion goes nowhere. This article tackles that problem head-on by proposing a 10-layer abstraction model for AI agent infrastructure—a shared conceptual map for navigating the crowded landscape.

The piece opens with a telling scenario: a Runtime tool is introduced in a team review, but as the conversation unfolds it becomes clear the tool is also handling Orchestration and Memory concerns. This kind of category confusion is not a personal failing—it reflects a broader truth that the AI agent ecosystem has not yet settled on a common vocabulary. Even major frameworks like LangChain, AutoGen, CrewAI, and Semantic Kernel tend to describe themselves in terms that span multiple layers, making apples-to-apples comparison genuinely hard.

The proposed 10-layer model starts from the lowest level—model inference and raw LLM calls—and builds upward through tool invocation, memory management, inter-agent communication, orchestration logic, evaluation and observability, and finally deployment and serving infrastructure. By treating each layer as a distinct concern, teams can make more precise statements: this SDK is a thin wrapper around the Runtime layer; that framework takes on responsibilities well above Orchestration. The framework makes implicit architectural decisions explicit, which is where its practical value lies.

The Model Context Protocol (MCP), published by Anthropic and increasingly adopted across the ecosystem, maps naturally onto this kind of layered thinking. MCP standardizes how tools, resources, and prompts are surfaced to an agent, which places it primarily at the Tool Use and Context Injection layers. The fact that LangChain, LlamaIndex, and Semantic Kernel have all moved to add MCP support suggests the industry is beginning to converge on inter-layer interface standards—a process that layered architecture maps can help accelerate by giving teams a shared language for discussing what fits where.

From a practical standpoint, a map like this has uses beyond SDK selection. It can structure internal design reviews for agent infrastructure, provide a neutral evaluation grid when assessing vendor proposals, and help onboard new engineers by giving them a mental model before they dive into any specific toolchain. Frameworks that market themselves as "all-in-one" solutions are particularly worth scrutinizing through this lens: which layers are genuinely strong, and which are thin abstractions that will need supplementing?

The broader implication is that as agentic systems grow more complex—spanning multiple models, tool ecosystems, memory backends, and deployment environments—architectural clarity becomes a competitive advantage in itself. Teams that can precisely diagnose where a tool fits, and where gaps remain, are better positioned to build systems that remain maintainable as the ecosystem evolves. This article offers one well-considered attempt at that kind of clarity.