Faced with hundreds of millions of lines of aging code, Siemens is deploying an interconnected network of AI agents to overhaul the legacy software that keeps factories running, energy grids powered, and transportation networks moving.
The, called Knowledge Fabric, was detailed in a Google Cloud blog. It said Siemens faced codebases spanning hundreds of millions of lines, developed over more than a decade. It described Siemens software as supporting factories, energy grids and transportation networks.
While the scale of the codebase was one part of the issue, Google said critical context was spread across code repositories, Jira tickets, Confluence pages and scanned PDF manuals from the early 2000s.
Siemens also had to work within industrial quality, compliance and product-lifecycle requirements that can extend across 15 to 20 years of operation.
Building an automated lifecycle system
Siemens and Google Cloud built Knowledge Fabric as a software development lifecycle automation system using knowledge graphs on Spanner Graph, the Google Agent Development Kit, Gemini API, Agent Platform, Gemini CLI and Anthropic Claude Code.
“By ingesting the entire software ecosystem into an intelligent agentic system equipped with custom knowledge graphs, we aren’t just helping developers optimize their development time; we are enabling autonomous agents to reason across the past to build the future,” said Franz Menzl, senior vice president, product creation excellence at Siemens.
“This is about freeing engineers from repetitive work so they can focus on higher-value problem solving.”
Moving beyond standard RAG
According to Google Cloud, Knowledge Fabric maps relationships between assets instead of treating every file as isolated text.
It said the system uses graph queries, vector search and full-text search to connect code snippets to requirements in design documents, identify dependencies and answer impact questions such as which functions would need to change if logic in a control-panel feature were updated.
The graph structure addresses a limitation Google identified in standard retrieval-augmented generation. “We realized that standard RAG (retrieval-augmented generation) wasn’t enough,” said Agata Gołębiowska, technical lead at Google Cloud.
Code has structure, Gołębiowska said: classes belong to files, files belong to modules and those relationships can be lost when a codebase is flattened into a vector database.
Dividing work among specialized agents
The workflow then breaks large modernization requests into smaller agent-led tasks. The blog post outlined five agent roles: a search agent that explores the code graph, a user-story agent that gathers requirements, an architecture-impact agent that checks likely side effects, a task-breakdown agent that turns analysis into smaller work items and a coding agent that implements specific tasks.
Human review remains part of each stage, according to Google. The risk is clear from the company’s own post, which states that hallucinated or unvalidated changes are “operationally unacceptable” in industrial-grade software.
NIST’s OT security guidance makes the same risk environment clear, describing operational technology as systems that interact with the physical environment and have distinct performance, reliability and safety requirements.
Evaluating pilot results and operational impact
Siemens used Knowledge Fabric in a production pilot to migrate legacy control panels to modern web-based interfaces. Google said dependency analysis for a new feature had previously taken senior engineers several days across codebases and legacy documentation. With Knowledge Fabric, the same work now takes “far less time,” according to the post.
Google said the pilot reduced overall coding effort while preserving system integrity and industrial quality standards, but it did not publish a percentage reduction, number of engineers involved, cost impact, testing process, security-review method or deployment timeline.
Navigating technical debt and industrial governance
The broader technical-debt context is well documented. The Consortium for Information & Software Quality estimated that poor software quality cost the U.S. at least $2.41 trillion in 2022, with accumulated technical debt reaching about $1.52 trillion. CISQ also identified technical debt as “the biggest obstacle to making any changes to existing code bases.”
For industrial companies, that obstacle is not only a developer-productivity problem. Software changes can affect products, maintenance obligations, compliance evidence and operational reliability. Siemens’ own Industrial AI material describes its approach as connecting, contextualizing and using shop-floor and enterprise data across the industrial value chain.
The governance question follows from the same structure. NIST’s AI risk guidance says trustworthy AI systems depend on characteristics including validity, reliability, safety, security, resilience, accountability, transparency, explainability, interpretability, privacy and fairness with harmful bias managed.