Building Automation System (BAS) evolution.
What you haven’t seen is the decades-long, silent revolution happening inside the wall-mounted controllers that make it all possible. This isn’t just a story of smarter thermostats; it’s a fundamental shift from isolated, pre-programmed devices to a network of intelligent, cloud-augmented “actors.”
A recent article, “Why the BMS Industry Should Embrace Python for Control Programming,” serves as the perfect trigger for this discussion. For those of us who’ve been in the trenches of building automation, this question isn’t merely about a new programming language. It behooves a much larger, more profound question: Are we ready to move true Artificial Intelligence to the edge, and in doing so, finally break free from the architectural constraints that have limited our industry for decades?
The Legacy Era: From Pneumatics to Limited Microprocessors
To understand the future, we must first appreciate the past. The journey begins long before the digital microcontroller.
- The 1960s-1970s: The Pneumatic Age. Before DDC, there was air. Building control was executed by sophisticated networks of pneumatic thermostats and actuators running on compressed air. The “logic” was mechanical, using bellows, levers, and air pressure relationships. These systems were robust, simple, and inherently safe, but they were imprecise, slow to respond, and incapable of any complex computation or centralized management. The “programming” was done by a technician physically adjusting springs and nozzles.
The original Direct Digital Control (DDC) controllers of the 1980s and 90s were the digital marvels that replaced pneumatics, but they were still severely handicapped by the standards of today.
- Hardware Constraints: We’re talking 8-bit microprocessors with clock speeds in the single-digit MHz and memory measured in kilobytes. There was no room for bloated operating systems or high-level interpreters.
- The Rise of Proprietary Control Languages (PCLs): To squeeze logic into these tiny devices, manufacturers developed their own highly efficient, proprietary control languages. Think of names like Siemens’ PPCL, Honeywell’s Caret, or Johnson Controls’ Metasys MEC/MPC. These were often simple, line-oriented languages perfect for executing basic control sequences:textSET TEMP = AI1 IF TEMP > 75 THEN SET AO1 = 100
- The “Weeds” We Got Lost In: As noted in Toby’s article about actors, which sets the scene for this evolution, this era “limited the industry” by creating a paradigm of isolated, hierarchical “actors.” Each controller had a single, rigid job. The system architecture was a reflection of the hardware’s limitations—a tightly coupled but brittle tree of devices. The deep knowledge required to navigate these proprietary ecosystems created immense vendor lock-in and stifled innovation. We were experts in the “weeds” of a particular system, not in the overarching “garden” of building intelligence.
The Timeline of Unshackling: A Technological Evolution
The journey from then to now is a story of progressive unshackling, driven by Moore’s Law and network connectivity.
| Era | Hardware | Software & Connectivity | Industry Impact |
|---|---|---|---|
| 1960-1970s | Compressed Air, Brass, & Rubber | Pneumatic Logic | Reliable but imprecise. No data, no centralized management, no complex sequences. |
| 1980s-90s | 8/16-bit microprocessors, KBs of RAM | Proprietary Control Languages (PCLs), Serial Networks (BACnet MS/TP, LON) | Closed systems, vendor lock-in, focus on basic control loops. The Digital Revolution begins. |
| 2000-2010 | 32-bit processors, MBs of RAM | Standardized Protocols (BACnet/IP), Embedded Linux, Graphical Programming (e.g., Niagara Framework) | Linux emerges as the silent disruptor, providing a stable, open OS foundation. Interoperability begins, IT/OT convergence starts. |
| 2010-2020 | Powerful ARM CPUs, GBs of RAM | Cloud Platforms, RESTful APIs, Open Source (Python, JavaScript) | Data becomes as valuable as control. Cloud analytics for fault detection & diagnostics (FDD). |
| 2020-Present | Raspberry Pi-class SBCs, AI Accelerators | Linux, Containerization (Docker), Python, TinyML, Small Language Models (SLMs) | The “AI at the Edge” era begins. Linux + Python forms the core stack for intelligent, containerized applications. |
The Modern Inflection Point: The Linux-Python Stack and Pushing Intelligence to the Edge
This brings us to today’s pivotal moment. The call to embrace Python is only half the story; it is enabled by its powerful partner: Linux.
- Linux: The Foundational Operating System: While Python is the language of innovation, Linux is the bedrock that makes it possible. Its arrival in the 2000s, often embedded within broader frameworks, marked a quiet revolution. Linux provided a stable, secure, and open-source operating system that could manage hardware resources, networking, and security in a way that bare-metal firmware or RTOS (Real-Time Operating Systems) never could. It turned a proprietary controller into a standardized, network-connected computer.
- Python & Linux: A Symbiotic Relationship: The question of moving to Python is symbolic, but its power is unlocked by Linux.
- Linux provides the ecosystem: It offers the filesystem, process management, and network stack that Python and its vast library ecosystem depend on.
- Python provides the accessibility: It allows for rapid development of complex logic, data analysis, and now, AI models. A data scientist can build a predictive algorithm in Python on a laptop, and thanks to Linux, it can run with minimal modification on a Raspberry Pi at the edge.
This symbiotic relationship is what finally breaks the proprietary lock. We are no longer programming a controller; we are developing software for a Linux-based industrial appliance.
- AI at the Edge: The Ultimate Synthesis: This is where the stack converges. We are now “pushing the power to the edge.” A modern controller runs Linux, which orchestrates containerized Python applications executing Small Language Models (SLMs). This enables:
- True Predictive Control: A VAV box controller can learn the thermal personality of its specific zone and pre-emptively adjust airflow.
- Natural Language Commands: A maintenance tech could ask a controller, “What’s wrong with you?” and the local SLM could interpret the query and respond in plain English.
- Resilient Operation: Intelligence survives a network outage. The system degrades gracefully, making local, smart decisions instead of failing entirely.
The Expert’s Viewpoint: A New Paradigm for Building “Actors”
The scene set by the article on actors is more relevant than ever, but we must now rewrite the script. The old DDC systems were like actors who could only recite one line, in one play, for one director. The new paradigm, enabled by the Linux-Python stack, creates actors who are improvisational masters.
They understand the context of their immediate environment, can collaborate intelligently with their fellow “actors,” and can take creative direction from both the cloud and local occupants. This is what forward-thinking experts are already doing, pushing the boundaries of what’s possible at the network’s edge.
The challenge for our industry is no longer technical feasibility; it’s architectural and cultural. We must design systems where security is paramount, where data flows are bidirectional and semantic, and where the business model shifts from selling proprietary hardware to providing continuous value through intelligent software.
The journey from the pneumatic thermostat to the cloud-conscious, Linux-powered AI edge device has been a long one. The question raised by the article—Should we move to Python?—is incomplete without acknowledging its partner. The real shift is to the Linux-Python stack. This stack is the gateway to moving AI to the edge, and its adoption will define the next decade. The building of the future won’t just be automated; it will be intelligent, adaptive, and finally, truly responsive to the humans and the planet it serves.
Doug Migliori, MBA provides this perspective
Digital Twin Consulting / Decision Intelligence / Event Driven Systems / AI-powered Airports
I’m increasingly skeptical of “L” words — programming and markup languages were designed for human readability, not agent-native communication. Their syntax and semantics serve human authorship, compiled or interpreted for machine execution. But AI agents don’t need linguistic scaffolding. They operate more efficiently through machine-optimized data streams: metadata, embeddings, ontological bindings, and capability graphs. In agent ecosystems, language becomes overhead. The future isn’t about expressing logic in human-readable form — it’s about orchestrating behavior through structured, interoperable signals that transcend syntax entirely.
Moreover, embedding a unified metadata model within both physical and digital twins allows for precise alignment between simulated and real-world behavior. By declaratively encoding the systemic and physical characteristics of building components, metadata ensures that virtual simulations mirror the operational logic, constraints, and capabilities of actual equipment. This harmonization not only enables predictive modeling and scenario orchestration, but also fosters interoperability — between digital twins, across heterogeneous physical systems, and at the critical interface between physical and digital realms. Metadata becomes the lingua franca that bridges domains, vendors, and runtimes.
I advocate for metadata-driven control logic that can be either compiled or interpreted by runtime engines built in the most appropriate language for their host environment — whether that’s C++ for real-time microcontroller systems or Python for cloud orchestration. This approach ensures performance, portability, and semantic clarity across deployment tiers.
