Picture a pit mine in the Pilbara at three in the morning. A driverless haul truck the size of a two-story house rounds a bend on a haul road, sensors reading dust, grade, and the position of a water cart two hundred meters ahead. No one is in the cab. No one is watching a dashboard in real time. The truck makes a decision, executes it, and logs it. Then it does that again, several thousand times before sunrise.
Caterpillar has been running systems like this for years. Now, according to TechCrunch, the company is applying what it learned from automating mining to how it deploys AI. I want to sit with that for a moment, because from where I sit as someone who spends her days thinking about agent architecture, this is a more interesting transfer of knowledge than it first appears.
Autonomy Is an Operations Problem Wearing a Research Costume
Most of the agent work I read about optimizes for capability. Can the model plan? Can it use tools? Can it recover from a failed subtask? These are good questions and largely the wrong ones for anyone trying to actually field a system.
Mining automation had to answer a different set of questions, and it had to answer them in an environment where a mistake is measured in tons and lives. Not “can the machine drive itself” but: who is accountable when it doesn’t? What is the degraded mode? How do you certify a fleet where every unit has slightly different wear characteristics and firmware? How do you retrain the workforce whose job just changed shape? How do you prove to a regulator, in writing, that the thing is safe?
Those are operations questions, and they are the ones that determine whether an autonomous system survives contact with the world. An organization that has already been through that gauntlet in one domain carries something that no amount of benchmark performance provides: institutional muscle memory for the boring parts.
The Parts That Actually Transfer
Here is what I would expect to carry over from heavy equipment autonomy to enterprise AI deployment, based on the structural similarities between the two problems:
- Staged autonomy rather than a switch. Mining automation did not go from manual to driverless. It went through remote operation, then supervised autonomy, then semi-autonomous fleets. Each stage generated the telemetry that justified the next. Agent deployment tends to skip this and jump straight to “give it write access.”
- Telemetry as a first-class product. An autonomous haul fleet produces obsessive amounts of data about its own behavior, because you cannot manage what you cannot observe. Most agent deployments I have looked at have observability bolted on after the first incident.
- Designing for the human in the loop who is bored. Automation does not eliminate operators; it changes what they do and, critically, makes their attention intermittent. Mining learned hard lessons about handoffs to humans who have not been paying close attention. Agent systems are about to learn the same ones.
- Fleet heterogeneity as the default assumption. No two machines in a real fleet are identical. No two model versions, prompt configurations, or tool permissions in a real agent deployment are either. Systems that assume uniformity break in ways that are miserable to debug.
Why This Lands Now
The rest of this week’s news makes the timing legible. An Anthropic researcher gave a peek at self-improving AI. Warp introduced a system it describes as an out-of-the-box software factory for AI development. Hugging Face is selling a $399 open source robot duck called Microduck. The capability curve and the tooling curve are both moving fast, and the accessibility floor is dropping to the price of a decent mechanical keyboard.
What has not moved nearly as fast is the deployment discipline. We are getting very good at making agents that can do things and remaining fairly bad at making organizations that can safely run them. That gap is where a company like Caterpillar has an unusual advantage, and it is not a modeling advantage. It is a process one.
The Uncomfortable Implication
If the hard part of agentic AI turns out to be operational rather than architectural, then the organizations best positioned to field it are not necessarily the ones building the models. They are the ones who have already shipped autonomy into a physical environment with real consequences and lived through the aftermath.
That reframes the competitive picture. A century-old equipment manufacturer with a decade of autonomous fleet operations behind it may understand something about deploying agents that a frontier lab does not, precisely because the lab has never had to explain a failure to a mine safety inspector.
I would rather read that company’s incident post-mortems than most of the papers on my desk right now. The pit does not grade on a curve.
🕒 Published: