The Robot's Agentic Moment Hasn't Arrived Yet

I spent today visiting several robotics and humanoid companies, and came away with a clearer sense of where the real bottleneck sits. A few observations.

Hardware has already started commoditizing. A company may still pull off a genuine hardware breakthrough, but competitors — especially in China — will close the gap quickly. What remains defensible is not the machine itself but the ability to manufacture it at scale, efficiently. When the business has to grow, capacity becomes the moat.

Deployment in the real world is still the hard part. Standing up a humanoid for an actual use case typically takes three to nine months. The work is stubbornly bespoke: gathering data, training the vision-language-action model, and getting the robot's AI to act on it in concert. To teach a robot something as mundane as picking a parcel off a warehouse rack, you might need around 200 lines of high-quality training data — roughly a skilled engineer's week. There is no turnkey solution on the market.

The mature use cases aren't the ones you'd expect. Entertainment — a robot that dances — and education — research labs in universities, demonstrations that inspire students — are the furthest along. Logistics and manufacturing, the use cases everyone is betting on commercially, are still developing with a huge amount of manual work needed during implementation.

The recurring theme is that the gap is no longer in the body. It's in the software. Despite enormous investment in all the physical-AI models (like VLA), most robotics work today is still the painstaking stitching together of those models manually to make something function in a real environment.

Here's the analogy I keep returning to. Large language models existed for a while before they became genuinely useful — what changed was the agentic moment, the shift that turned a clever text predictor into a system that could plan, act, self-correct, and learn. Robotics hasn't had that moment yet. The physical-AI models exist, but nothing has fused them into a universal, self-learning, self-correcting system. The honest benchmark for maturity is simple: a robot should not take longer than a human to learn a given task.

So one of two futures seems likely.

In the first, the agentic moment arrives. A general-purpose brain emerges, and humanoids become quick to deploy — learning new tasks at human speed or better.

In the second, that universal brain turns out to be impossible. The industry reverts to a specialized approach: providers build tightly customized hardware-plus-software systems for each use case. And if that's where we land, it raises a sharper question — do we even need the humanoid form at all? If you're building bespoke anyway, you build whatever shape best fits the task.

Which future we get depends entirely on whether robotics finds its agentic moment. Today, it hasn't.