The real AGI-pill is specialized robots

Most of the excitement around physical AI is about general-purpose robots that can be dropped into any factory or home and take on any task a human can do. I think this misses half of the picture. If the goal is abundance for everyone, a world where the cost of everything approaches the cost of the atoms and energy required to make it, you need to squeeze costs to the minimum, and in manufacturing that has historically meant specialization. AI also makes specialized machines cheap: agents design them, and general-purpose robots build them.

Here’s how I expect automation to evolve:

• Automation has been mostly hand-programmed specialized machines, only worth it for frequent, stable tasks.
• In the short term, AI makes general-purpose robots viable. They take on tasks too variable or too infrequent to justify a specialized machine.
• In the long run, agents make specialized machines cheap to design, and general-purpose robots make them cheap to build. General-purpose robots take on each new task, and once it becomes frequent enough, a specialized machine takes it over. I expect a Cambrian explosion of specialized machines.

I start by laying out why industrial automation is valuable. Then I introduce robot learning and AI agents, discuss how they lower the costs of automation, compare general-purpose robots with specialized machines, and finally explain why this leads to the future above.

Contents

• Industrial automation
• Robot learning
• Agents
• Costs of automation
• General vs specialized
• Going forward
• Links

Industrial automation

It’s common to take for granted that industrial automation is great and desirable. In this section I’m breaking down why, in order to establish common ground to later discuss robot learning and general-purpose robots in the industrial context.

The most straightforward case for automation is cost reduction: replacing a recurring cost (one or more workers paid by the hour) with an upfront investment in the machines and engineering for the project. Companies often screen automation projects using payback time: the upfront investment divided by the annual savings it generates. In the simplest case, those savings are the displaced labor costs minus new expenses for maintenance, energy, supervision and other operating costs.

Payback time depends heavily on utilization: the more hours the machine runs, the more its cost gets spread out and the lower the effective cost per unit produced. A machine that saves costs at full capacity may be a poor investment if demand is volatile or it sits idle for long chunks of time. Utilization will be important throughout the blog post: we’ll use it to discuss specialized vs general-purpose machinery: a specialized machine can be utilized intensively when its task is frequent and stable, while a more general machine can preserve utilization by being reassigned when demand or products change.

Payback time is important, but stronger cases appear when automation relaxes a binding constraint on production, rather than simply cutting costs. Its value then is in the production it unlocks, not the wages it displaces.

Some industrial processes benefit from superhuman performance. A machine can apply forces and move at speeds beyond the human range. It can consistently position a component within tens of microns without its motion drifting from fatigue or distraction. Repeatability leads to fewer defects, and therefore reduced rework time and scrap. Other processes include heavy, dirty or hazardous tasks that put human health at risk. When semiconductor fabs moved to 300mm wafers, a loaded carrier crossed the weight a person can safely lift and move all day, which made automating transport a physical necessity rather than a cost decision.¹

Sometimes a company can’t find workers. For example, in Italy in 2025, companies planned 13,600 entries for specialist electric welders and expected 80% of them to be hard to fill.² If production is bottlenecked on welding, the company is forced to either overpay, forgo growth or change its processes. In that case, the value of a welding robot is not in cutting labor costs, but in removing the bottleneck that prevents the company from scaling.

Automation can also amplify workers instead of replacing them. If we consider a job as a bunch of tasks, machines can take over repetitive and dull tasks, potentially freeing workers to focus on tasks where human judgment creates the most value. For example, professional welders might program and supervise a fleet of welding robots, exerting their specialized knowledge through the machines and thus multiplying their output and reach.

Some human characteristics make us less well-suited to industrial processes than machines: we fatigue, change jobs, take time off, and get sick. But people also have qualities that no machine can currently match: we are generally intelligent, can learn and adapt to multiple tasks, exercise judgment when situations are ambiguous, respond to contingencies, coordinate with other people, and creatively improve the processes around us.

Human flexibility also makes us useful across a changing mix of tasks, even when no single task would justify a dedicated machine. This is why today’s factories are often made of specialized machines glued together by human workers. This glue is made of physical tasks and cognitive tasks. Physical tasks include moving parts from one machine to the next, changeovers when the product changes, clearing jams, replacing worn tools, pulling defective parts from the line. Each of these tasks is often either very hard for hand-programmed automation or too infrequent to justify a specialized machine from a utilization standpoint. Cognitive tasks include deciding what to automate, designing and programming the cells, scheduling production, diagnosing failures, dealing with suppliers.

Robot learning

The holy grail of robot learning is to enable robots with any embodiment, operating in any environment, to carry out useful real-world tasks from natural language instructions or short demonstrations. We are not there yet. We have foundation models that can perform simple, short-horizon tasks after a single demonstration,³ and others that can be fine-tuned to reach reliable performance on a specific task across unseen environments.⁴ Progress is fast; I expect the previous sentence to get outdated quickly.

Public demonstrations mostly establish capability, not deployability. A successful demo shows that a robot can perform a task, usually under controlled conditions. A deployed automation system must perform it thousands of times, while recovering safely from unexpected failures. Yet those unexpected failures are exactly the cases for which data is hardest to collect. Repeating a task the policy has already mastered adds little. Model capabilities improve with novelty, including corrective interventions and failures.⁵ There’s a mismatch between environments where data on long-tail edge cases can be collected most effectively, and environments where a robot can be deployed productively.

This is a reason many companies focus on home environments. As Sunday Robotics puts it: “homes are spaces of infinite variation, long tails and constant change.”⁴ Homes are a more natural source of novelty than constrained industrial settings. They demand open-world generalization, making them well aligned with the broader research objective of building general robotics foundation models.⁶

Many industrial applications are narrower and more repetitive, but have stricter requirements along different dimensions: reliability, safety, consistent cycle times, and auditability. Hype aside, manufacturers don’t care whether a cell uses hand-programmed automation, a learned policy for one subtask, or an end-to-end robotics foundation model. They care about whether the system achieves their goals. Because failures can interrupt production or create scrap, one practical near-term architecture limits learned policies to variable steps inside otherwise deterministic processes, for example cable insertion and soldering in electronics manufacturing.⁷ Other deployments post-train foundation models for specific tasks, for example order packaging in e-commerce.⁸

Robot learning also enables automation of new classes of tasks that hand-programmed automation has struggled to tackle, for example cable manipulation, clothes folding, adaptive welding. Basically anything with lots of variability and feedback dependence, which includes many of the physical tasks in the glue between specialized machines.

Agents

By agents I mean systems built on frontier language models that plan, use tools and act over long horizons, like the coding agents that already write a large share of software. It’s not hard to imagine that those agents will also be able to take over the cognitive tasks in the glue between specialized machines, and eventually design and run entire factories.

Pieces of this are starting to come together:

• Frontier models can control robots directly, with no robot-specific training, and it’s quite effective given their increasingly general capabilities. The main limitation is latency: models take seconds per call, too slow to run a robot in real time.^9101112
• MicroFactory sells cells with robot arms, cameras and swappable grippers, and gives an agent the tools to do everything in the cell, replacing the system integrator. You describe the task, the agent designs the grippers and tooling for 3D printing and writes the computer vision code that runs the robot. The agent controls the robot through the code it writes rather than directly, which gets around the latency problem from the previous bullet.¹³
• SendCutSend automates quoting, programming the machines, scheduling and billing.¹⁴ SendCutSend predates coding agents, which make something like this much cheaper to replicate in other industries.
• Saturn Robotics is building software that stores 3D models and repair histories of its own in-house machines so breakdowns get fixed faster, and will simulate production to plan how many machines and people it needs, which AI coding made tractable for a small team.¹⁵
• Agents are getting good with CAD. OpenAI reports GPT-6 Astra at 95.9% geometric overlap on BenchCAD,¹⁶ but on RealCADBench, built from industrial parts, the six frontier models tested range from 28% to 54% solid IoU.¹⁷
• Project Vend started in 2025 with Claude running a vending machine in Anthropic’s office, with mixed results, and got much better in its second phase, mostly thanks to better tools and procedures.¹⁸ Andon Labs, Anthropic’s partner on the project, has since opened a store in San Francisco run by an agent that handles procurement, staff scheduling, the website, banking and deals with suppliers, while humans do the physical work. It’s not profitable yet.¹⁹

Costs of automation

The costs of automation projects go beyond the price of the machines. I’ll group them in five categories: designing the process, building the cell, programming, validation, and maintenance and retasking.

A lot of work goes into designing how the machine fits in the broader production process: how parts arrive and leave, what tolerances are acceptable, how defective parts are detected, etc. The cell then has to be built, typically by a system integrator; this often includes safety and certification work. The machine’s task must be programmed and validated to meet the requirements. After installation, it needs to be maintained, with spare parts and technical support readily available. Factory operators or external system integrators may also need to retask the machine as business conditions change, paying again for design, programming and validation.

An automation project has to be validated not only for whether it can perform the task, but also for whether it remains reliable under production conditions and fails safely and predictably. During operation, failures are often expensive: the machine may need to be repaired, production may stop, parts may need to be scrapped or reworked, downstream processes may be starved of inputs, etc.

Robot learning and agents act on different categories. This is how they change them in the short/medium term:

Agents will automate factory design work like they are automating coding: you describe the cell you need and the agent designs it in days instead of months. Building it out requires physical work though. Agents can source components and guide humans through the installation, but someone still has to assemble and wire the cell. A related cost is safety certification, which is partly institutional and therefore might resist automation for longer.

Both forces reduce the cost of programming the cell. For example, today to make a welding robot productive you need both specialized welding knowledge and specialized robot programming knowledge. In the future you might not need the latter: you can just tell or show the robot how you want the piece to be welded. Robot learning approaches this by letting a skilled worker program a robot with language, or through demonstrations and corrections. Agents synthesize programs from language instructions, and additionally help with parts that will likely always be code, such as PLC programs.

Lower programming costs are not free, in particular robot learning comes with an increase in validation burden. Hand-programmed automation is also not guaranteed to work under every condition and relies on reliability metrics, especially if there is a vision component, but it’s generally more bounded and reproducible. The possible failure modes of a learned policy are harder to enumerate. Moreover, as the policy improves, failures become harder to find: long-tail failure modes require, by definition, a growing number of trials as they get rarer, which gets expensive quickly when evaluating in the real world.²⁰⁵ It’s not just about failure rate and how long the robot can operate continuously, it’s also about how the robot fails. Operators therefore need to spend time making sure the robot behaves as they expect, collecting corrective demonstrations and removing potential sources of costly failures.

Lower programming costs also mean cheaper retasking. A hand-programmed automation project is a bet on the forecasted demand for the products currently manufactured by a company. If the actual demand differs or other conditions change, the company might need to drop or change the product or its production process. This makes the investment risky, since the robot might be unfit for the new setup. Both robot learning and agents controlling robots allow the same robot to be reassigned to new tasks, lowering the risk of the hardware investment and amortizing its cost across all the tasks it performs during its lifetime. This raises utilization, which can shorten payback times and improve economics for tasks performed less frequently for which the investment could not be justified. This fits high-mix low-volume production particularly well. Retasking through software is limited by the robot’s hardware envelope: the tasks its hardware can physically perform. Agents can design new hardware for the tasks outside it, but for now installing the hardware and repairing the machines is still something humans have to do.

Together, agents and robot learning lower automation costs enough to carry many projects over the investability threshold. In the last section we discuss how general-purpose robots will help bridge the gap with what’s left: physical work like building cells and maintenance.

General vs specialized

Automation hardware sits on a spectrum from specialized machinery to general-purpose robots: precision semiconductor equipment, systems of conveyors, feeders and pneumatic mechanisms, single 6-axis arms with a specific tool as end effector (e.g. a screwdriver),²¹ arms with parallel jaw grippers, systems with multiple arms, mobile robots, and humanoids with dexterous hands. Robot learning shows up at both ends of the spectrum. At the general end, it’s what makes general-purpose robots possible. At the specialized end, it handles the few variable steps inside otherwise deterministic machines, like a welding robot with sensors that adapts to each piece. Agents make the specialized end cheaper to design and operate.

A more general embodiment supports more tasks: a fixed arm will never carry a box across the factory floor, a mobile robot on wheels will never walk up a flight of stairs. Its larger hardware envelope lets retasking through software reach more tasks, an optionality that is worth more, in terms of utilization and shorter payback time, the more the process being automated is uncertain. The optimal robot is the cheapest embodiment whose hardware envelope covers the expected tasks at the required performance.

Specialization, on the other hand, buys throughput and other performance characteristics that a general-purpose hardware envelope does not optimize for, by stripping out everything the task does not need. A press or a feeder can run thousands of cycles an hour for years with a handful of actuators. Fewer actuators also mean fewer parts that wear out and less energy per unit produced. For some tasks there’s no alternative to specialization: when the binding constraint on production is speed or micron-level precision, a general-purpose robot can’t do the job. Specialized machines are also easier to validate: they run deterministic code, written by hand today and increasingly by agents, and their failure modes are more bounded than those of a learned policy. The catch is the upfront investment: someone has to design, build, program and validate the machine, and that’s only worth it for tasks frequent and stable enough to keep it busy.

The strongest argument for humanoids is the optionality above: a worker that can take on any task at any time. Other arguments come up often, here’s how I see them:

• Data transfer from humans. Humanoids share our body, so they should benefit the most from foundation models trained on human data. Probably true, but human data seems to help other embodiments too. Dyna pre-trained a model only on human video, and the more video it saw, the better it predicted the actions of bimanual arms with parallel jaw grippers.²² Physical Intelligence found a similar transfer in a model controlling an arm with a parallel jaw gripper.²³
• Economies of scale. General-purpose robots can be produced in higher volume than specialized machines, and will therefore ride down the learning curve faster. But 50% to 70% of a humanoid’s BOM is actuators,²⁴ and cheaper actuators make simpler robots cheaper too. A Unitree G1 has 23 to 29 actuators without counting the hands, a 6-axis arm has 6. Simpler embodiments also cover more tasks than they seem, e.g. a power gripping tool instead of a dexterous hand.²⁵
• Brownfield compatibility. A humanoid can work in an existing factory without changes. This helps in the short term, but factories built around people aren’t optimal, and sometimes the human body is the very bottleneck automation removes.
• Footprint. Legs let a robot carry heavy loads without a large base, so a humanoid needs less floor space than a wheeled mobile manipulator.²⁶ Floor space doesn’t matter for many tasks though.

Going forward

Today’s factories are often made of specialized machines glued together by human workers. We’re moving towards being able to automate this glue, with general-purpose robots taking over the physical tasks and agents taking over the cognitive tasks. We can imagine the general-purpose robots acting as an extension of the agents, carrying out tasks in the real world. Future factories will be systems of:

• agents running the factory finances and production floor, identifying bottlenecks and designing solutions to them, often in the form of specialized machines
• general-purpose robots building new machines, maintaining equipment and handling exceptions
• specialized machines, some with learned steps, replacing general-purpose robots whenever a task becomes frequent and stable enough to justify them from a utilization perspective
• humans overseeing everything and setting the general direction for the factory

The ratio between general-purpose robots and specialized machines will evolve over time. General-purpose robots will take on new and changing tasks. As bottlenecks are identified, specialized machines are designed to tackle them, freeing general-purpose robots to work on the next task. Repetition continually creates opportunities for specialization, while novelty continually recreates a residual role for general-purpose robots. The bottleneck keeps moving as the production process improves and the cost per unit drops toward the cost of the atoms and energy that go into the product.

This kind of loop appeared in many industries, for instance semiconductor fabs automated wafer transport one bottleneck at a time, from carts pushed by people, to automated guided vehicles, to vehicles running on ceiling rails between tools.¹ Until now the loop has been driven by humans, and only when volumes justified the engineering. Soon it will be driven by agents acting in the physical world through general-purpose robots. The loop will turn faster, and across more industries, including in high-mix low-volume factories.

For example, take welding: a 6-axis welding robot with sensors can see the piece and adapt to deformations, so an experienced welder only needs to design the path once on a sample piece. The next bottlenecks are designing the path for each new piece and loading and unloading the fixture. A sufficiently capable agent can generate the path from the CAD of the piece, and a general-purpose robot can load the fixture. If a piece is ordered in high enough volume, the agent can go further and optimize for throughput, for instance by designing dedicated fixtures and a conveyor so that the robot can weld continuously and without stopping, or by adding parallel welding robots. The general-purpose robots build those machines, which are specialized to the task at hand, and move on; the bottleneck moves to other parts of the manufacturing process. Robot learning shows up twice here: in the welding robot adapting to each piece, and in the general-purpose robots building the line around it.

Building machines is a great fit for general-purpose robots: it’s exactly the kind of variable, infrequent task that would hardly justify a specialized machine. Robots can also take machines apart when a product is discontinued, and agents can reuse the components in the next design. This lowers the upfront investment required for specialized machines, thus lowering the utilization threshold needed for a given payback time.

Cheap intelligence does not make specialization go away:

• Some tasks need performance general-purpose robots can’t reach.
• If every company can buy the same general-purpose robots, one way to preserve a company’s competitive advantage is building a specialized production process around the product. Unlike software, a factory takes time, energy and atoms to replicate somewhere else, so the advantage holds even when anyone can copy its design.
• By Jevons paradox, abundance creates the volume specialization needs.
• When designing and building machines is almost free, what’s left is the cost of atoms and energy. A machine stripped down to one task uses the least of both, while a general-purpose robot carries actuators and mass the task doesn’t need.

Cheap intelligence lowers the cost of operating general-purpose robots, but also the cost of designing and deploying specialized machinery: agents design the machines, and general-purpose robots build them. You can’t predict the future of automation by considering only the first effect. I expect a Cambrian explosion of specialized machines, built for tasks that today are too infrequent to justify one.

This post started as one on how Europe could remain sovereign in the AI race. I split off a post on automation in European industry, then split this one off from that. If those sound interesting, follow me on x dot com.

Links