Agent Product徹底比較 — 33製品を自律度×提供形態のカオスマップで整理するThis article maps 33 AI agent products on a chaos map using autonomy level and…

The rapid proliferation of AI agent products has made it increasingly difficult for teams to compare options on any consistent basis, and a new entry in the "AI Agent Reference Architecture" series on Zenn addresses that by plotting 33 agent products onto a single chaos map. Rather than ranking tools by brand recognition or feature checklists, the article organizes the market along two axes—autonomy level and delivery model—to give readers a structural framework for selecting products beyond brand names.

The chaos map format is a familiar convention in Japanese technology writing, where a crowded field of products is arranged on a two-dimensional grid so the overall landscape becomes legible at a glance. Here, one axis measures how autonomously a product operates, ranging from tools that require step-by-step human direction to systems that plan and execute multi-step tasks with minimal supervision. The other axis captures the delivery model, distinguishing, for example, hosted software-as-a-service offerings from libraries, frameworks, or self-managed runtimes that developers embed in their own stack.

By separating these two dimensions, the article makes a point that is easy to miss when products are discussed by name alone: two tools that sound similar in marketing terms can occupy very different positions once autonomy and delivery are considered independently. A highly autonomous cloud service and a low-autonomy developer library may both be described as "AI agents," yet they imply different integration work, operational responsibilities, and risk profiles.

This comparison sits within a larger series that argues for understanding AI agents by layer rather than by tool name. The parent article frames itself as a map for people who want to understand AI agents, and companion pieces extend the same structural lens to other parts of the stack. One examines execution engines—the control layer that coordinates work in what the author calls the multi-agent era—while another compares how agents handle memory, dividing the topic into built-in memory, external memory, and a newer, more speculative layer the author describes with the metaphor of "dreaming." A separate, more implementation-focused note observes that Ollama is weak at parallelism, reflecting the series' interest in practical runtime constraints.

The reference-architecture approach the series adopts mirrors a broader industry shift. As agent tooling has expanded, frameworks such as LangChain, LlamaIndex, Microsoft's AutoGen, CrewAI, and orchestration layers like LangGraph have proliferated alongside vendor products from OpenAI, Anthropic, Google, and others. Standardization efforts, including Anthropic's Model Context Protocol for connecting agents to external tools and data, point to the same underlying need: a shared way to reason about components rather than an endless catalog of brand names. Viewing the field by layer—models, memory, execution, tools, and the agent surface itself—helps clarify where any given product actually operates.

For readers, the practical value of the chaos map is likely in narrowing choices before deep evaluation. Placing a candidate on the autonomy axis clarifies how much human oversight it expects, which in turn affects testing, guardrails, and accountability. Placing it on the delivery axis clarifies who owns operation and scaling. Because the map appears to be a snapshot of a fast-moving market, positions are best read as an orientation aid rather than a fixed verdict; product capabilities and categories shift frequently, and any single grid necessarily simplifies nuanced differences.

The article does not claim to be an exhaustive benchmark, and its 33-product scope is a curated sample rather than the entire market. Its contribution is conceptual: offering a repeatable way to situate agent products so that comparisons rest on structure rather than hype. For teams weighing where to invest, that framing—paired with the series' companion analyses of execution and memory—may prove more durable than any individual tool recommendation, precisely because the vocabulary it builds is intended to outlast the specific products plotted on the chart.