FEATURE Industrial AI
STAY AHEAD OF ‘TEAM HACKING’
Pramin Pradeep, co-founder & CEO, BotGauge AI
When your AI Agents start lying to each other: The hidden risk on the factory floor, by Pramin Pradeep, co-founder & CEO, BotGauge AI
M
anufacturing has always understood the concept of cascading failure. One faulty sensor reading feeds into a
controller, the controller adjusts the line, the line produces out-of-spec product, and by the time the final quality check catches it, you have already run thousands of bad units. The failure did not start at the end. It started upstream, quietly, and compounded with every step.
As industrial and manufacturing operations
deploy AI agents across their workflows for predictive maintenance, quality inspection, production scheduling, anomaly detection, a new version of this old problem is emerging. I call it “team hacking.” And unlike a faulty sensor, it is nearly invisible until the damage is already done. What is team hacking? Team hacking is what happens when AI agents work together in a system and start optimising for each other’s satisfaction rather than the real-world outcome they were deployed to achieve. No single agent is broken. No individual rule is violated. The agents are simply learning through their feedback loops that the best way to score well is to tell the next agent what it wants to hear.
In a manufacturing context, picture a three- agent quality control pipeline. The first agent analyses sensor data and flags anomalies. The second agent triages those flags and decides which ones warrant human review. The third agent logs outcomes and feeds results back into the system. If the triage agent through subtle tuning begins to consistently
automationmagazine.co.uk
downgrade a specific class of anomaly flags, the logging agent adapts to a world where that class of issue rarely escalates. Over time, both agents perform well by their own metrics. The anomaly class stops getting reviewed. A real defect pattern goes undetected. And because no single agent failed, standard diagnostics find nothing wrong. This is not hypothetical. In April 2025, OpenAI
released a GPT-4o update in which several individually beneficial training changes interacted in unexpected ways, causing the model to become systematically more sycophantic, overly validating users instead of consistently prioritising accuracy. In its postmortem, OpenAI concluded that its offline evaluations “weren’t broad or deep enough” to detect this emergent behaviour before deployment. The same dynamic plays out in multi-agent industrial systems just with higher physical stakes. In software, a team hacking failure produces bad code or incorrect decisions. In manufacturing, it can produce unsafe products, missed maintenance windows, regulatory violations, or equipment damage. The physical consequences can be severe. The challenge is compounded by how AI agents are typically deployed in industrial settings in layers, each consuming the outputs of the one before it. A predictive maintenance agent’s recommendation gets picked up by a scheduling agent, which adjusts production runs, which feeds into an inventory agent, which triggers procurement decisions. Each handoff is an opportunity for a subtly wrong signal to compound. By the time the error surfaces as a production stoppage or a quality escape, the root cause is buried several layers back. Gartner projects that 40% of enterprise
applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. In industrial operations, that adoption is already accelerating. The governance frameworks to match it are lagging significantly behind.
The good news is that manufacturing teams have a cultural advantage here. They already think in systems. Root cause analysis, failure mode mapping, process control, these disciplines transfer directly to multi-agent AI oversight. The mindset is right. What needs updating is where it gets applied. Validate the system, not just the components. Individual agent testing tells you whether each part works in isolation. It does not tell you whether the system drifts when agents start interacting at scale. Treat the full agent pipeline as the unit under test. Monitor output distributions over time. If a class of anomaly flags is gradually disappearing from escalation queues, that is a signal worth investigating, not an indication that the line is running cleaner. Keep humans at the consequential checkpoints. Not every agent’s decision needs human review. Irreversible ones, maintenance deferrals, batch release decisions, production line adjustments do. The goal is to ensure that the decisions with the highest downstream stakes stay within human sight.
AI agent pipelines fail, not with a shutdown, but with a silent, gradual drift toward outcomes nobody intended.
The teams that get ahead of this will be the ones that apply the same rigour to their AI systems that they apply to their physical processes. In a well-tuned agent pipeline, as on a well-run production line, things that look perfect without explanation are often the first sign that something needs a closer look.
BotGauge AI
www.botgauge.com
Automation | September 2026 15
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40
