By Shannon Holgate, Co-Founder at Hypership
At the frontier of AI, companies are shipping more code generated by AI than typed by people. That shift took about two years. Why software first? The usual answer is abundant code online and the technical talent clustered around it.
There’s another, more boring answer: software already has the machine that tells you when you’re wrong. A compiler says no. A test fails. The change appears on screen in red and green. That machine gave agents a feedback loop. AI got good at software because it could be told it was wrong, cheaply, thousands of times a day.
Agents get better, faster, where the work can answer back. The same AI that is great at writing code can be surprisingly poor at directing you to a nearby car wash. This is jagged intelligence.
READ MORE: Why the best careers keep stretching you
If coding intelligence is Everest, directions intelligence is Slieve Donard: still a mountain, but not reliable enough for the backstreets in your village. Jagged intelligence becomes expensive and dangerous when its output enters the physical world.
Unlike software, the physical world cannot afford thousands of cheap failed attempts, and it does not hand back test results so readily. In one lab’s published results for its own robotics model, task success ranges from roughly one in three to nine in ten, depending on the task.
So agents will land next in the domains that build repeated, low-cost feedback loops. Some industries already generate the proof. Modern welding machines can record data from each weld. A pipe run either holds pressure or it does not. That proof exists at the point of work, but it is not wired into a system that can learn from it.
Physical industries look like the last place agents will land. On this argument, they might not be. I see what the other end looks like most weeks. A tool is signed out on paper, then moved between sites until the record no longer matches reality. A materials sheet is filled from memory at the end of the working day.
Without that record, neither the agent nor the owner can know what happened, let alone whether it was right. One construction platform agreed to spend $845m this summer acquiring the ability to see a site and check it against the plan.
READ MORE: Developing people for the roles that don’t exist yet
Field service software in preview now turns a technician’s spoken update into suggested changes and asks them to confirm those changes before they update the record. Both create checked evidence at the point of work.
That is how the physical world begins to answer back. None of this removes the human sign-off. In radiographic weld inspection, AI can flag possible defects, but the inspector still reviews the result and signs. The responsibility remains theirs.
The inspector receives more evidence and signs with greater confidence. The responsibility remains theirs. What I don't know yet is whether industries will wire up their own proof because it pays, or only when someone else's agent turns up and demands it.
Read the Summer 2026 edition free online →
Stay connected with NI's tech community:
