Key ideas
- Pay goes to whatever limits the whole system, so when agents make a kind of work cheap, its pay falls.
- What stays scarce is judgment, taste, someone willing to be responsible, physical work on site, trust with a buyer, and ownership.
- A person who runs agents well does the work of several people, and he expects that multiple to keep growing.
- If workers' share of income keeps falling, the other share, ownership, is the one to hold.
- For most of these roles there's a fast way in: a short certificate, public bug bounties, or selling a technical product at a smaller company.
- Build a service skill and a name first, then a product, because a product built now may be out of date by the time it's done.
Nick Saraev’s test for which careers will still pay is one question: what will still be scarce once agents do most of the work? Pay goes to the bottleneck, so he looks for work that needs judgment, taste, a person to be responsible, or hands on a job site. That gives him six: forward deployed engineer, electrician, AI red teamer, AI media producer, enterprise closer, and owner of a mostly-agent company. These are my notes on his 38-minute video, with his four premises, what each role does day to day and the first step he suggests. Salaries and stories are as he tells them; the research figures are checked against their sources.
Four premises
Agents work longer on their own. METR, a research group that measures AI agents, reported in March 2025 that the length of task agents can finish half the time had been doubling about every seven months since 2019. Saraev says today’s models already run for days inside the AI labs, and he expects the doubling to continue.
Pay moves to the bottleneck. Programming paid well because people wanted more software than programmers could write. Once a model can write it, he says, that pay falls. Doctors and lawyers are paid well for the same reason: there are always more patients and cases than people qualified to handle them.
The centaur. A human with AI is like a centaur: the person brings judgment, the agent brings speed and volume. He says agent hours at Anthropic already outnumber human hours by six or seven to one, so one person running agents well does the work of six or seven.
Owners win when labor is cheap. The Bureau of Labor Statistics put the labor share at 54.1% in the first quarter of 2026, its lowest since the series began in 1947 (a preliminary figure). He ties this to the argument that a government that earns its money from AI rather than from its people, like an oil state, has less reason to invest in them. That argument comes from Luke Drago and Rudolf Laine’s essay series The Intelligence Curse.
Forward deployed engineer
The job: go into a business, have the people who own each process walk you through it, and sort the tasks into ones an agent can do alone and ones that need a person working with an agent. Hand the first kind to agents outright. For the second, teach the staff to use agents as thinking partners. He says the first kind can give gains of thousands of percent and the second a few hundred.
In small businesses, under about a hundred staff, he starts with growth rather than back office, because that’s where they struggle. His shorthand for it is reach, acquisition, conversion and expansion. To map a process, he records or screenshots the person working (with their permission), has a model turn the log into written procedures, then automates them with browser or computer use.
His point about the day: building is no longer the hard part. Getting into businesses and getting them to explain how they work is. The tools are a coding agent such as Claude Code or Codex, plus voice dictation or a meeting transcriber feeding notes straight to the agent.
Demand is real. In October 2026 Anthropic committed $100 million to train 10,000 of these engineers by the end of 2027, through a program it calls Claude Frontier Academy. He puts median pay at $200,000 to $250,000.
Electrician
AI needs data centers, data centers need power, and no model can pull cable. The Department of Energy found data centers used 4.4% of US electricity in 2023 and projects 6.7% to 12% by 2028. The head of the Associated Builders and Contractors puts electrical work at about half of all labor on data center projects. Saraev expects supply of electricians to lag that demand into the 2030s, until robots can do the wiring.
The other reason it pays is responsibility. A licensed electrician signs off on the work and carries the liability, and that’s something an AI can’t take on yet.
His route in: an optional few months of pre-apprenticeship at a trade school, then three to five years as a paid apprentice before you’re licensed. Two faster ways onto data center sites are low-voltage work (fiber and network cabling, certified in days to weeks) and solar installation (about 40 to 60 hours of training). Rules vary by place, so check yours.
AI security and evals
As agents get access to email, money and sensitive permissions, every company handing over that access needs someone to show it’s safe. Red teamers attack models to find how they fail; evals score how well or how safely a model does a job.
You can start without a job or a degree. Anthropic’s model safety bug bounty, run on HackerOne, pays up to $35,000 for a new universal jailbreak (participants sign a non-disclosure agreement). He points to Pliny the Liberator on X as an example of the work. The methods he names are prompt injection (text an agent reads that it mistakes for instructions), role-play framing, and getting a model to leak secrets. These days, he says, it mostly means using models to attack models.
AI media producer
When anyone can produce anything for close to nothing, the scarce part moves from making to choosing what to make. That’s taste: picking, from a million possible outputs, the few people will want. He says costs are modest now (around ten cents a second of video) and heading toward the cost of electricity.
Upwork’s report on 2025 hiring found AI video generation and editing was its fastest-growing AI skill, up 329%. The tools he names are video and image models such as Seedance, Seedream, MiniMax and Kling, usually reached through a platform like Higgsfield that puts image, video and audio in one place. His own setup gives a coding agent API access to those models, so he directs the angles and styles and the agent runs the generations.
Enterprise closer
The title today is enterprise account executive: the person who closes multi-million-dollar deals with large companies. When the technical work is cheap, he argues, the constraint is convincing a big buyer to commit. He says Anthropic hires more for sales than for technical roles, and that much of it is in person, because a visit counts for more once pixels are free.
The route in is merit: technical sales at small software companies, meeting setting, or customer support, then larger deals. Pay is mostly on-target earnings, what you make if you hit your quota. Agents prune and qualify lead lists and handle the follow-up.
A one-person AI company
If labor’s share falls, the other share is ownership. Carta, which tracks startup ownership, found solo founders started 23.7% of new startups in 2019 and 36.3% in the first half of 2025.
He looks for work you do once that keeps paying: videos, educational material, a small tool for one industry, a paid newsletter, or an agency where agents deliver and the founder sells. His own day is the work that needs his judgment (filming, sales calls, decisions his team brings him), with agents doing the research and logistics.
His advice is not to start with a product. Build a service skill in whatever is scarce at the moment, such as forward deployed work, learn a few industries from the inside, then use that reputation to sell something more scalable later.
Using the test
The six roles will date. The question behind them won’t: for any job, ask which part stays scarce when agent labor is nearly free, and move toward that part. In his list it’s one of six things: judgment about what to automate, physical work with liability attached, proving agents are safe, taste, trust with a buyer, or owning the thing the agents work for.
Key terms
- Time horizon
- METR's measure of AI progress: how long a task takes a human expert, for tasks an AI agent completes half the time.
- Bottleneck
- The one step that limits how much the whole process can produce, and so the step that gets paid most.
- Centaur
- His picture of a person working with AI: the human supplies judgment, the agents supply the volume.
- Labor share
- The part of the economy's output paid to workers as compensation, as opposed to going to owners.
- Forward deployed engineer
- An engineer who works inside a customer's business, finds the work agents can do, and builds the systems to do it.
- Red teaming
- Attacking an AI system on purpose to find where it fails, through jailbreaks, prompt injection and similar methods.
- Eval
- A test that scores how well, or how safely, an AI does a job.
- Universal jailbreak
- A technique that reliably gets a model to break its rules whatever the request, the kind bug bounties pay most for.
Quotes from the talk
The bottleneck there is not actually doing the building.
One of the things that you cannot replace... is human responsibility.