Excel VBA × 生成AI API の事例を、AIエージェント4体に世界中から探させた話The author deployed four AI agents in parallel to scour the world for…

A developer who has spent a long-running blog series wiring Microsoft Excel's VBA macro language to generative AI APIs recently turned the tables on their own project, using four AI agents working in parallel to search the world for other people doing the same thing. The exercise is a small but telling example of how autonomous agents are increasingly being used not just to write code, but to conduct research about how software is actually used in the wild.

The premise of the series, according to the author, has been consistent: calling a generative AI API from inside Excel VBA should not stop at a single throwaway macro. Instead, the goal is to cultivate a collection of small tools, treated as a "group of instruments," that can be maintained and run in everyday business work. This framing matters because Excel remains one of the most widely deployed pieces of business software in the world, and VBA is still embedded in countless spreadsheets that handle real operational tasks. Bridging that legacy environment to modern large language models is a practical way to bring AI assistance to workflows that would otherwise be difficult to modernize.

The specific question that prompted the investigation was straightforward: how many people around the world are already doing this? Rather than manually combing through search engines, forums, and code repositories, the author appears to have delegated the discovery process to four AI agents running at the same time, each presumably tasked with scouring different sources or regions. Sharing what they found, and reflecting on why this kind of automation matters for daily business workflows, is the stated purpose of the piece.

Technically, calling a generative AI API from VBA usually involves issuing HTTP requests, often through objects such as MSXML2.XMLHTTP or WinHttpRequest, sending a JSON payload with a prompt and parameters, and parsing the JSON response back into worksheet cells or variables. Because VBA lacks native JSON handling, developers typically add helper functions or third-party parsers, and they must manage API keys, rate limits, and error handling carefully. These constraints are part of why building a reusable toolset, rather than one-off macros, is an appealing approach.

The article is filed under the Model Context Protocol category and carries tags for MCP and MCP server, which situates it within a broader industry conversation. MCP, introduced by Anthropic, is an open standard designed to give AI models a consistent way to connect to external tools, data sources, and services. In principle, an MCP server can expose capabilities that agents call in a structured way, and the protocol has been gaining adoption as vendors look for interoperable ways to let models act on real systems. The connection between an Excel-and-VBA toolset and MCP is a natural one: both are about turning AI from a chat window into something that manipulates concrete tools and data.

The use of multiple agents in parallel also reflects a wider trend. Frameworks and products for orchestrating several agents at once, sometimes called multi-agent systems, have proliferated alongside the rise of capable models from providers such as OpenAI, Anthropic, and Google. Coding-focused agents and research assistants can now browse, retrieve, and summarize information with limited human oversight, though their output still requires verification. Running four agents to survey global usage of a niche technique is a lightweight illustration of that capability, and it likely produces a broader sweep than a single searcher could manage in the same time.

For readers, the value of the piece is less in a definitive census of Excel-VBA-plus-AI practitioners and more in the demonstration of a method. It shows how a hobbyist or professional can combine a familiar tool like Excel, accessible generative AI APIs, and emerging agent orchestration to both build automation and study how widespread that automation has become. As always with agent-driven research, the findings should be treated as indicative rather than authoritative, since coverage depends on which sources the agents reached and how their results were interpreted. Still, the approach hints at how everyday developers may increasingly use AI to map, and then extend, the tools they already rely on.