Published Sep 9, 2026

In my opinion, the reconfigurability of a humanaoid robotic workforce has been underdiscussed. There is no precedent for being able to convert a million construction workers into a million doctors overnight.

For the entire history of labor markets, skills were scarce because training was expensive, knowledge was difficult to preserve, and each worker had limited attention. Agents and robotics invalidate all three of these constraints, as their capabilities will be able to be copied for free, distributed around the world instantly, and continually improved by the whole workforce.

Division of Labor

If you chart US jobs into the following two axes:

• Specialized/Routine: Does the job require advanced occupation-specific training or can it be learned quickly?
• Cognitive/Physical: Does the job primarily manipulate information/relationships or the world/atoms?

And consider the technological priors:

• Agents can automate cognitive work, where “agents” are multimodal models in harnesses that allow them to operate software, which may be used to setup custom software automations (codebases/workflows/schedules).
• Robotics can automate physical work, where “robotics” are general machines (humanoids driven by world models), which may be used to set up custom machine automations (gantries/arms/manufacturing).
• Skills represent some kind of task-specific training, which could be as simple as a markdown file or as complex as a world model fine-tuned on CNC configuration.
• Speed/accuracy/cost/availability/integration of agents and robotics will continue improving in a feedback loop for the foreseeable future.
• The government talks about AI regulation, but its practical stance to industry seems to be “let it rip”.
• Agents will be given authority to allocate resources, and will outperform humans at leveraging those resources.
• Historically, new work opportunities have appeared in prior technological revolutions.
• Adoption lags capability, due to demand/reliability/regulation/bureaucracy.

And the economic priors:

• Some things simply require humans.
• Marginal propensity to consume decreases with wealth and income.
• Wages have grown slower than productivity since the 1970s.
• Gains from productivity accumulate among capital owners in the absence of wage competition or collective bargaining.
• Capital and labor can substitute for each other.
• Labor compensation is a cost to firms.
• Businesses have financially demonstrated an insatiable hunger for automation, and they will not turn down free money.
• Agents will make markets more observable, allowing capital to identify and enter scarcity bottlenecks faster.

There follows some uncomfortable conclusions:

• Cognitive work is at high risk of humans becoming “meat proxies” that take observations and do tasks for machine decisions, since models can chip away at every part of a job that can be done on a computer (writing/coding/research/communication)
• Cognitive work is most exposed because agents using software tools have the fastest feedback loops and it is the juciest target, as it represents ~70% of income. This is unfortunate because routine cognitive work is typically a class gateway. Ironically, specialized cognitive fields have some resistance because they have inherent mechanisms to defend themselves, such as credentials, regulation, liability rules, and artificial scarcity.
• Physical work is safer for now because interacting with the real world is difficult, but companies will come for the remaining ~30% as returns improve. Much like full self driving, if it can be hill climbed, it will be. Capital will automate whichever physical work becomes financially viable as robotics costs decrease and capability increases.
• “Make more money” may be rational individual advice, but in order for the system to continue functioning, at some point, ownership has to broaden. This could look like employees receiving a larger share of company returns, or income becoming less dependent on employment, such as through UBI, a negative income tax, or dividends.

Stores of Value

Under this regime there are narrored sources of value capture:

• Scarcity/Ownership: Same as always! Creating/joining a first mover, doubling down on bets, being born lucky, or leveraging an information asymmetry are all good strategies for those with means to pursue them. For example, someone could own a factory and rent skilled machines to operate it, instead of having to develop their own robotics stack.
• Trust/Brand: Companies pay to have issues guaranteed to be solved by other entities, with risk pricing adjusted by the entities’ delivery histories. A skill provider could package suites of machine skills under a brand name that guarantees some level of accuracy/liability. Skills will need to be discovered, trained, curated, and updated.
• Access/Scale: Marketplaces are always a good business model, though as the effort cost of online price/quality/service-comparison decreases, they will now need to compete on things besides mindshare, such as service (Amazon’s easy returns), content (Netflix exclusive media), social (steam library/friends/reviews), or ecosystems (App/Play Store momentum/support). Skill marketplaces become possible once interchange formats standardize and skill providers begin to compete.
• Collective Bargaining: As long as some humans are needed, collective bargaining remains powerful. Blocking automation progress (the current US dockworker union strategy) is a poor long-term plan, as it increases capital’s incentive to automate human labor, and companies are generally not super interested in paying training data residuals. A better strategy is to bargain over future gains from automation, such as higher pay, fewer hours, equity, or perks.
• Revenue Proximity: In the short term, cognitive compensation will likely adjust to become more tightly tied to revenue generation, although when money flows start get controlled via machine-machine coordination, humans will not able to keep up. Skills that allow effective resource allocation will become more valuable as the amount of resources controlled by machines increases.

So what should you do about it? Generalists tend to outperform specialists in uncertain environments, so adapting quickly is itself a valuable trait. My aforementioned image annotation task went from impossible -> hard -> easy -> trivial in 20y. If any part of your career touches a computer, capital wants to automate it, and their swarms will compete at millions of tokens per second.

Good luck.