$399. That’s the price of Microduck, the open source robot Hugging Face is now selling — a small, cute, buyable machine that anyone with a credit card can put on a desk this week. I bring it up because on the same week that number crossed my feed, so did a headline from TechCrunch about Caterpillar bringing what it learned from automating mining to how it deploys AI. Two ends of the same spectrum: the cheapest possible robot and some of the largest machines humans have ever driven into a pit. And the interesting part is that the hard problem is identical at both scales.
It isn’t the model. It’s the deployment.
Why the mining comparison lands
I build bots for a living, mostly software agents that touch real systems — ticketing queues, deployment pipelines, internal APIs. The demo is never the hard part. I can get an agent doing something impressive in an afternoon. What eats months is everything after: what happens when the agent is wrong, who notices, how fast it stops, what it’s allowed to touch, and how you prove any of it to someone who is accountable for the outcome.
Mining automation is that problem with physics attached. When a machine that weighs more than a house makes a bad decision, you don’t file a bug and ship a patch on Friday. You build the operational envelope first and the autonomy second. So a company that spent years working out how to run automated equipment in an environment where mistakes are expensive and irreversible has a genuinely different instinct about AI than a company whose worst-case failure is a weird chatbot reply.
I don’t know the internals of what Caterpillar is doing — the news is the direction, not the architecture. But the direction itself is a useful signal for anyone building bots, and here’s what I’d take from it.
Four things heavy industry gets right that agent builders usually skip
1. The stop button is a feature, not an afterthought
In industrial automation, the ability to halt a machine is designed in from the first sketch, on its own path, independent of the smart part. Most agent frameworks I’ve used treat cancellation as a nice-to-have. If your bot can start a long chain of tool calls and you have no reliable way to interrupt it mid-chain, roll back partial writes, and see exactly where it stopped, you don’t have an autonomous system. You have a fast one.
2. Autonomy is graded, not binary
Heavy equipment didn’t jump from manual to driverless. It went through assist, then supervised remote operation, then narrow autonomy in constrained zones, then wider operation. Software people love to ship “the agent handles it now.” A better pattern is a ladder: the bot proposes, a human approves; the bot acts in a sandbox, a human reviews; the bot acts in production on a small class of tasks with logging; then you widen the class. Each rung earns the next one with evidence.
3. The environment is part of the design
Automated mining works partly because the site gets shaped to suit it — defined routes, known zones, predictable traffic rules. The equivalent for bots is unglamorous and enormously effective: give your agent clean, well-typed tools instead of a raw shell, narrow permissions instead of an admin token, structured data instead of scraped HTML. Making the world legible does more for reliability than a bigger model does.
4. Telemetry is the product
Industrial operators instrument everything, because you cannot manage a fleet you can’t see. Most bot projects I review log the final answer and nothing else. Log the plan, the tool calls, the inputs, the retries, the token spend, the latency, the human overrides. Overrides especially — they are your highest-signal dataset for what your bot is actually bad at.
What this means for the rest of us
The same week brought news of Warp shipping a system it describes as an out-of-the-box software factory for AI development, and an Anthropic researcher showing a look at self-improving AI. Both are pointed at capability. Caterpillar’s angle is pointed at deployment. My bet is that the second category ends up mattering more for the people reading this site, because capability is increasingly something you buy and deployment discipline is something you have to build.
Which brings me back to the duck. A $399 open source robot means the barrier to putting an autonomous thing into a real room is now a rounding error on a laptop. That’s great. It also means a lot of us are about to learn, at small scale and low stakes, the lessons that mining operations learned at large scale and high stakes. We can either rediscover them or read the notes.
Start with the stop button.
🕒 Published: