Robotics Is Having Its Agentic Moment

Everyone's watching humanoids. The real shift is in-context learning hitting robot arms.

Remember december 2025, when I started this report? That was the time software stopped being scripted automation and started being agents. Systems that perceive, reason, and act instead of blindly executing step 4 after step 3. The same shift is now hitting the physical world, and my take is that most people are watching the wrong thing. They’re watching humanoids do backflips on stage. The real story is happening somewhere much less photogenic.

The old paradigm: RPA in metal

Here’s the thing about industrial robots today: they’re basically RPA in metal. Hyper-specialized, pre-programmed for 1 task, 1 SKU, 1 fixed environment. The arm picks the box because the box is always the same box, in the same place, under the same lighting. Change anything, a new product shape, a new bin layout, and you’re back to weeks of costly reprogramming.

That’s exactly why automation stalls in high-mix environments. Fashion logistics, e-commerce fulfillment, anywhere the products change faster than the engineers can reprogram. The economics just don’t work. Sound familiar? It’s the same reason scripted software automation broke the moment a form field moved 10 pixels.

The shift: same recipe, new substrate

Now draw the mapping, because it’s almost embarrassingly clean.

A scripted workflow is to an agent what a pre-programmed robot is to a generalist robot. In software, you went from “if this, then that” pipelines to systems that figure out the steps themselves. In robotics, you go from teach-pendant choreography to machines that work out how to grasp the thing in front of them.

The enabling stack is the same too. Foundation models for reasoning. Vision models for perception. And the big one: in-context learning instead of retraining. You don’t fine-tune an agent every time a new edge case shows up. You give it context and it adapts. A generalist robot handles a new SKU the exact same way. Show it, don’t retrain it.

Which brings us to the killer feature, and it’s identical in both worlds: adaptation without redeployment. That’s the entire value proposition of agents in software, and it’s the entire value proposition of generalist robots in the physical world. The robot that meets a never-seen-before object and thinks okay, irregular shape, soft material, grip from the side is doing exactly what your coding agent does when it hits an undocumented API.

The proof point

This isn’t just a thesis. The InCoRo paper, co-authored by researchers from Theker, Meta AI, and ServiceNow, demonstrates an LLM-driven control loop that uses in-context learning to pilot industrial robot arms in dynamic environments, with no task-specific retraining, reaching 83.2% success rates on SCARA units in changing conditions. An LLM controller, a scene understanding unit, a feedback loop. That’s literally the agentic recipe, pointed at a physical arm instead of a browser.

The contrarian bit

One more thing, because the hype machine will tell you otherwise. The agentic moment in software didn’t arrive through AGI demos. It arrived through boring, deployed workflows: support tickets, code review, data entry. Expect the exact same pattern in robotics. Not humanoids waving on a keynote stage, but adaptive arms quietly running in warehouses, handling SKU number 4,000 without anyone filing a reprogramming ticket.

The open question

Here’s what I keep coming back to. In software, the hard part of agents turned out to be reliability and feedback loops, not reasoning. The models could think; the systems couldn’t be trusted to run unsupervised. Is that also the bottleneck for embodied agents? And what does the equivalent of evals look like for a robot, when a failed test case isn’t a wrong answer but a dropped package, or worse?