When agents clash: Stopping algorithmic civil war on the plant floor

Consider autonomous software algorithms making millisecond decisions in complete isolation from one another. Welcome to the multi-agent pileup.

What you’ll learn:

  • Enterprises face a severe operational hazard: the "arbitration vacuum."
  • On a factory floor, conflicting commands cause thermal cycling, cavitation, hydraulic water hammer, and catastrophic mechanical breakdown.
  • Naive AI architects are proposing an "anti-pattern." But using a probabilistic LLM to supervise other probabilistic agents is an unmitigated disaster.

Editor's note: This is the third article of a new five-part series by Colin Masson of ARC Advisory Group where he's taking a no-nonsense look at how the latest wave of AI collides with the physical laws of the plant floor.

He’s stripping away the hyperbole about AI, deconstructing the vendor marketing narratives, and providing plant managers, operations executives, and enterprise architects—part of our core audience—with actionable blueprints for governed and profitable implementations.

See Part 1: Why billion-dollar AI model claims fall apart on the factory floor

See Part 2: The ‘AI-native’ mirage: Why pragmatic wrapping beats marketing washing


For most of my career, I assumed the hardest coordination challenge on an industrial site was getting the mechanical maintenance crew and the process engineering team to agree on a scheduled shutdown window.

In the physical world, when human teams disagree, they debate it over coffee in the control room, weigh throughput against asset wear, and eventually compromise on a sensible operating setpoint.

See also: Give operators the authority to stop flawed AI workflows

Now, imagine removing the coffee, the control room conversation, and the human common sense—and replacing them with autonomous software algorithms making millisecond decisions in complete isolation from one another. Welcome to the multi-agent collision.

The brownfield 'arbitration vacuum'

Following major enterprise rollouts—such as Salesforce Agentforce, SAP Joule, and autonomous agent frameworks from major automation vendors—enterprises face a severe operational hazard. We’ll call it the "arbitration vacuum."

Consider this real-world operational clash:

  • The reliability mandate (OT): An asset performance management agent monitors vibration telemetry on a primary boiler feed pump. It detects early bearing micro-friction and initiates an automated script to de-rate the pump by 30% to preserve equipment life until the scheduled weekend shift.
  • The commercial mandate (supply chain/ERP): Simultaneously, an ERP supply chain fulfillment agent parses a sudden spot-market demand surge carrying severe contractual penalties for late delivery. It updates the execution schedule and commands the line to overdrive that exact same feed pump to 105% capacity.

Without deterministic meta-orchestration, these two software agents wage an invisible, high-speed war across your supervisory control layers. The APM agent throttles the pump back, the ERP agent commands it forward, and the control loop oscillates.

In pure software, an infinite loop causes high CPU usage. On a factory floor, conflicting commands cause thermal cycling, cavitation, hydraulic water hammer, and catastrophic mechanical breakdown.

Why ‘market liability’ fails in operations

Silicon Valley platform leaders have argued that market discipline and legal liability are sufficient guardrails for autonomous systems: If an agent causes damage, post-hoc legal liability will hold the vendor accountable.

See also: Manufacturing’s hardest hire is the person running the automation

In consumer software, digital media, or enterprise SaaS, post-hoc liability works reasonably well. If a recommendation algorithm glitches or a chatbot hallucinates an inaccurate product description, lawyers negotiate settlements, PR teams issue apologies, and developers push an overnight software patch. On the factory floor, however, a post-hoc legal settlement is completely useless.

A lawsuit does not rebuild a compromised chemical distillation column. A legal settlement does not reverse an environmental toxic release. And commercial liability does not restore power to a regional electrical grid knocked offline by an oscillating control loop. Physical operations demand hard-coded, deterministic prevention, not courtroom compensation.

Architectural anti-pattern: The probabilistic ‘meta-supervisor’

To solve this coordination breakdown, naive AI architects are currently proposing an architectural anti-pattern: deploying a large "meta-supervisor LLM" to sit above the operational agents and arbitrate their disputes conversationally.

In mission-critical industrial operations, using a probabilistic LLM to supervise other probabilistic agents is an unmitigated disaster.

In consumer software, digital media, or enterprise SaaS, post-hoc liability works reasonably well. On the factory floor, however, a post-hoc legal settlement is completely useless.

If Agent A (probabilistic) and Agent B (probabilistic) are arbitrated by Supervisor Agent C (also probabilistic), you do not achieve governance. You achieve compounded hallucinations, non-deterministic drift, and runaway token burn.

See also: Relay research finds critical gaps in factory-floor information-sharing, leading to failures

The meta-orchestrator cannot be another conversational chatbot. It must be a hardcore deterministic state machine—such as AWS Step Functions paired with Bedrock AgentCore, or certified IEC 61131-3 logic on an edge controller. Deterministic state machines execute hardcoded priority rules:

  • Safety and environmental interlocks always override commercial throughput.
  • Asset reliability envelopes hold absolute veto power over schedule acceleration.

If an automated resolution exceeds programmatic financial or operational tolerances, the transaction trips an immediate Autonomy Revocation Protocol, freezes the asset in its last known safe state, and escalates out-of-band to a human engineer.

The economics of agentic work: Cloud Tokenpocalypse vs. Edge CapEx

Beyond operational collisions, continuous reasoning agents introduce a severe balance sheet risk: The Tokenpocalypse.

An autonomous agent running extended perception-reasoning loops across plant software consumes orders of magnitude more tokens than a standard single-turn chatbot.

If an enterprise tethers continuous reasoning agents to cloud-metered endpoints to monitor operational lines, software expense transitions from a predictable software subscription into a volatile utility bill that scales unpredictably with plant activity.

See also: Navigating financial anxiety around paying the bill for industrial AI

Furthermore, cloud reasoning encounters the immutable laws of network physics: latency. A cloud-hosted reasoning agent communicating across wide-area networks experiences variable round-trip times ranging from several seconds to several minutes per multi-step workflow. While a four-minute delay is acceptable for drafting a quarterly demand forecast, it is entirely unacceptable for supervising an active manufacturing process operating on millisecond control loops.

An autonomous agent running extended perception-reasoning loops across plant software consumes orders of magnitude more tokens than a standard single-turn chatbot.

To defeat the Tokenpocalypse and resolve the latency penalty, pacesetting industrial organizations execute an industrial edge escape:

  • Fixed capital investment (CapEx): Deploying high-density edge silicon—such as on-premises industrial PCs, edge appliances, or localized AI workstations (e.g., NVIDIA RTX Spark superchip appliances featuring 128GB of unified memory)—capitalizes inference as a fixed plant asset.
  • Zero marginal token cost: Running open-weight, quantized models (such as Llama 3.3, Qwen 2.5, or DeepSeek R1 Distill) locally on edge iron drops the operational cost per reasoning token to $0.00.
  • Air-gapped operational sovereignty: Response times are deterministic (sub-second), and if the plant loses wide-area network connectivity, production lines continue operating safely without interruption.

Centralized cloud platforms remain valuable for cross-site supply chain orchestration and periodic engineering design synthesis.

But when it comes to the plant floor, plant managers must own their runtimes: Deploy unmetered edge iron, lock in deterministic response times, and keep operational control safely inside the plant fence.

About the Author

Colin Masson

Colin Masson

Colin Masson is director of research for industrial AI at ARC Advisory Group and is a leading voice on the application of AI and advanced analytics in the industrial sector.

With more than 40 years of experience at the forefront of manufacturing transformation, he provides strategic guidance to both technology suppliers and end-users on their journey toward intelligent, autonomous operations.

His research covers a wide range of topics, including industrial AI, machine learning, digital transformation, industrial IoT, and the critical role of modern data architectures like the industrial data fabric. He is a recognized expert on the convergence of IT, OT, and ET.

Sign up for our eNewsletters
Get the latest news and updates