The 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:

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.

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.

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.

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:

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