Siemensが「象を切り分ける」手法で産業用ソフトウェア開発のエージェント型ワークフローを推進 How Siemens "slices the elephant," advancing agentic workflows for industrial software development

For a company like Siemens, software is not a peripheral concern — it is the nervous system of factories, power grids, and transportation infrastructure across the globe. As one of the largest industrial technology conglomerates in the world, Siemens operates at a scale where legacy software debt is not a theoretical worry but a concrete engineering challenge faced every day. The company has now partnered with Google Cloud to address that challenge using Gemini-powered agentic workflows, organized around a principle they call "slicing the elephant."

The metaphor comes from the familiar adage: how do you eat an elephant? One bite at a time. Applied to software modernization, the idea is to decompose a massive, seemingly intractable legacy codebase into discrete units that AI agents can analyze, transform, and validate incrementally. This stands in sharp contrast to the big-bang rewrite, a strategy that has burned enormous resources and failed spectacularly at companies ranging from startups to Fortune 500s. By working in bounded, verifiable slices, Siemens reduces risk while maintaining continuity of the industrial systems their customers depend on.

The technical foundation here is the agentic workflow model. Unlike a conventional AI assistant that responds to individual prompts in isolation, an agentic architecture deploys multiple specialized agents that coordinate to complete long-horizon, multi-step tasks autonomously. In Siemens' case, that means individual agents may handle distinct functions — code parsing, intent inference, test generation, documentation synthesis, refactoring suggestions — all wired together into a coherent pipeline. The human engineer shifts from author to reviewer, which can meaningfully increase throughput on large modernization programs without sacrificing quality control.

The problem Siemens is tackling has deep roots. Industrial software in manufacturing, utilities, and transportation frequently includes code written in COBOL, Fortran, or early C — systems dating to the 1970s and 1980s that remain in active production use decades later. The original developers are long gone. Documentation is sparse or nonexistent. The institutional knowledge of why certain design decisions were made is often irretrievably lost. Any tool that can plausibly reconstruct programmer intent from the code itself, and propose safe migration paths, represents a genuine advance in the state of the art.

This positions Siemens' initiative in an interesting place relative to the broader competitive landscape. Microsoft's GitHub Copilot and Amazon Q Developer have made AI-assisted coding mainstream, and both have invested in enterprise-grade transformation features. But those tools primarily augment individual developer actions — autocomplete scaled up. The agentic paradigm Siemens is exploring goes further: it hands full sub-tasks to autonomous agents that can execute, test, and validate their own outputs before surfacing results to humans. That orchestration capability, rather than raw model quality, may be the defining differentiator as enterprise AI matures.

Google Cloud's infrastructure appears well-suited to the task. Vertex AI and the Agent Development Kit (ADK) provide the scaffolding for multi-agent systems, and Gemini's strong code understanding capabilities make it a natural fit for the kind of legacy code analysis Siemens requires. The collaboration also underscores Google Cloud's broader strategic shift: positioning Gemini less as a standalone model and more as the engine of an enterprise agent platform.

Siemens itself has been building toward this moment. Its Xcelerator platform consolidates an industrial software portfolio spanning digital twins, simulation, and AI-driven design. Embedding agentic AI into the software development lifecycle is a logical extension of that strategy — useful not only for modernizing internal tools but for delivering faster, more reliable software to Siemens' own industrial customers.

If the approach proves replicable at scale, it could serve as a reference architecture for other large industrials confronting similar mountains of technical debt. The elephant, it turns out, can be eaten — as long as you're willing to pick up the knife and start slicing.