Pieter Van Schalkwyk

CEO at XMPRO

Mark O’Neill‘s recent post on AI agents struck a nerve with me. He points to OpenAI’s AgentKitas a prime example of this trend: a drag-and-drop workflow builder with RPA-style node connections that calls itself an “AI Agent builder” simply because one of the nodes might call an LLM.

We’re witnessing widespread “agent-washing” where vendors rebrand deterministic workflow builders as AI agents, and this matters enormously in industrial environments where decisions affect physical equipment, production schedules, and human safety.

The Workflow-Agent Confusion

Let’s be clear about what we’re seeing in the market today. A workflow is a predetermined sequence of steps where an LLM might be one of many nodes, and while customers can visualize using these drag-and-drop tools that demo well, calling this an “agent” misleads organizations about what they’re actually implementing. At XMPro we call this DataStreams with “Embedded AI”. They provide specific enhanced workflow capabilities, but they are not “agents” with “agency”.

Gartner defines AI agents as “autonomous or semiautonomous software entities that use AI techniques to perceive, make decisions, take actions, and achieve goals in their digital or physical environments.” The key word is autonomous. True agents don’t just follow your flowchart but observe their environment, reason about what they perceive, plan their approach, and take action while adapting to changing conditions and learning from outcomes.

The Carroll Framework: Separating Real Agents from Pretenders

Earlier this year, I developed what I call the Carroll Industrial AI Agent Framework, building on Michael Carroll‘s excellent work on “agent-washing” and Arthur Kordon’s foundational axioms for industrial AI. The framework asks three fundamental questions that most marketed “agents” fail to answer:

• Can it learn and evolve? A true agent improves through experience without constant external updates, showing measurable improvement in decision quality over time while adapting its approaches based on outcomes. This continuous evolution demonstrates the agent’s ability to build genuine understanding rather than simply accumulating data.
• Can it predict with causal understanding? The agent must grasp not only what happened but why it happened and what could happen next, understanding causality rather than just correlation. This enables proactive decision-making rather than reactive responses.
• Can it reason autonomously? The agent must make decisions independently within defined boundaries based on principled reasoning, not just execute predetermined logic paths.

Most “agents” being marketed today fail all three tests. They also fail what I call the “Cognitive Architecture Test” from my recent article on brain-inspired AI agents: do they have the specialized cognitive modules (memory systems, world models, reasoning pathways) working together in a coherent architecture that mirrors how biological intelligence actually works? A recent landmark survey paper, “Brain-Inspired Modular AI Agents,”authored by researchers from 20 top institutions including Stanford, Yale, and DeepMind, provides a comprehensive framework for this approach.

True intelligent agents require these specific cognitive modules working together, needing memory systems that preserve experience, world models that understand causality, and reasoning capabilities that adapt to new situations. Adding a chat interface to your workflow builder doesn’t create these capabilities.

Why This Matters in Industrial Contexts

In industrial operations, we deal with emergent processes where the next best step emerges based on real-time conditions. An experienced reliability engineer doesn’t follow a flowchart when diagnosing complex equipment behavior but observes patterns, reflects on similar past situations, considers multiple factors, plans an approach, and adapts based on results.

This differs fundamentally from content-based workflow agents that excel at document generation and predefined task sequences. Those have value, but they can’t replicate the judgment that keeps operations running smoothly when experienced operators retire and take decades of decision-making expertise with them.

The difference isn’t just technical but about impact on business performance. Workflow automation focuses on efficiency (doing things right), while true cognitive agents focus on effectiveness (doing the right things). Improving efficiency in the wrong direction just accelerates movement toward undesired outcomes, much like running faster in the wrong direction in a race.

The Path Forward

We need to address this agent-washing trend with practical steps that separate genuine capability from marketing hype.

• Be prescriptive about definitions. Use Gartner’s definition and insist on demonstrations of autonomous, long-running, tool-calling agents that can perceive, decide, and act independently. Don’t accept workflow builders rebranded as agents just because they include an LLM node somewhere in the flow.
• Look for the ORPA pattern (Observe, Reflect, Plan, Act). True agents don’t just process inputs and produce outputs but engage in a cognitive cycle that mirrors how human experts make decisions, continuously building understanding through observation and reflection.
• Evaluate against the Carroll Framework questions. Does the system demonstrate learning, causal reasoning, and autonomous decision-making? Does it have the cognitive architecture with specialized modules working together? If not, it’s probably workflow automation with better marketing.
• Start with appropriate complexity. Not every process needs a cognitive agent, as simple workflow agents and RPA have their place for routine, well-defined tasks. But when you need systems that can adapt to changing conditions and learn from operational outcomes, demand true agency rather than accepting rebranded automation tools.

The promise of AI agents is real. We have a front row seat watching leading companies use true agents to gain competitive advantage in complex operational environments. These systems transform how organizations handle intricate operational challenges while preserving decades of retiring workforce expertise and enabling decision-making at machine speed.

However, realizing this potential requires us to stop accepting rebranded workflow tools as agents and instead demand the autonomous, adaptive, learning systems the term actually describes. Without this clarity, widespread agent-washing will erode industry confidence in technology that could genuinely transform industrial operations.

Pieter van Schalkwyk is the CEO of XMPro, specializing in industrial AI agent orchestration and governance. Drawing on 30+ years of experience in industrial automation, he helps organizations implement practical AI solutions that deliver measurable business outcomes while ensuring responsible AI deployment at scale.

About XMPro: We help industrial companies automate complex operational decisions. Our cognitive agents learn from your experts and keep improving, ensuring consistent operations even as your workforce changes.

Our GitHub Repo has more technical information. You can also contact me or Gavin Green for more information.

Read more on MAGS at The Digital Engineer