Hire the AI First: Andrew Ng on Agentic Workflows and the Scarcity of Judgment

In the same week, two people at Andrew Ng's shop asked him for resources. The first wanted more headcount to push a project forward. Ng said no. Days later someone else came with a similar job and asked for budget to hire an AI to do it. That one he approved. Recalling the pair of conversations on *No Priors* with Sarah Guo and Elad Gil, he treats the instinct itself as the point: before you staff a task with humans, check whether an AI can take it. That reflex, he argues, is now a form of competitiveness.

Guo introduces him as one of the godfathers of the AI wave: co-founder of Google Brain, Coursera, and the venture studio AI Fund, coiner of "agentic AI," and a recent addition to Amazon's board. A decade earlier he was among the first people who convinced her that deep learning was the future. He starts not with the hiring story but with where capability growth comes from. There is still some juice in scale, he says, though it is getting brutally hard. Society's picture of progress has been skewed by a handful of companies with formidable PR, which made "just add compute" feel like the only vector. He wants several: agentic workflows, multimodal models, concrete applications, and wild cards such as whether diffusion models, mostly used for images, can generate text too.

He coined *agentic AI* over a teammate's objection: "Andrew, the world does not need you to make up another term." He did it anyway, because he was tired of people spending their energy on whether a system "is an agent." Agency is a spectrum. At one end are highly autonomous systems that plan, reason in multiple steps, and do a lot on their own; at the other, a prompt and a reply. Call the whole range agentic and get back to building. The word stuck harder than he expected. Marketers later used it as a sticker on everything in sight. Hype, he says, went nearly vertical. Real business progress has been growing fast too, just not as fast as the sticker.

Asked what still blocks true agents in applications, Ng grants the missing pieces. Computer use kind of works and often does not. Guardrails and evals are a huge problem: how do you evaluate these systems quickly and then actually *drive* the evals? Components have room to improve. The single biggest barrier he sees, though, is talent. Watching many teams build agents, the market's main differentiator is whether they know how to run a systematic error-analysis process with evals. At any moment a disciplined team can say what is working, what is not, and what to improve. Less experienced teams try things more randomly, and it simply takes a long time.

A lot of the work that could be automated through agentic workflows still needs that disciplined engineering process, and the skills, and maybe the tooling, are not widely in place. Could AI automate the process of building the agents themselves? Much of it, he says, requires ingesting external knowledge locked in people's heads. Until there are AI avatars that can interview employees doing the work, and better visual AI that can look at a monitor, he expects a year or two more of heavy lifting by human engineers and product managers. The context is proprietary. It is not sitting in the internet pre-training set or in a manual you can scrape.

The example he gives is next-generation robotic process automation. A customer emails a document; you convert it to text; you search the web for a compliance check so you are not dealing with a banned vendor; you look up pricing in a database and save it somewhere else. Multi-step, almost never correct on the first pass. Then the real questions start. Is a wrong invoice date a problem? Is routing a message to the wrong person for verification an incident? Is bothering the CEO to confirm invoices too often a crisis, or does the CEO not mind? Those judgments live in the people who run the business. Building the workflow still means sitting with them and deciding which errors matter.

On the spectrum of agentic AI, the strongest working examples Ng has seen are coding agents. He splits current economic value into two obvious buckets. One is answering people's questions, where ChatGPT had takeoff velocity. The other is coding agents. His personal favorite at the time is Claude Code: it plans how to build software, makes a checklist, and walks through it. The ability to plan and then execute multiple steps, in a system people actually use, is among the most autonomous agents in production. Computer-use demos that shop or browse for you can look magical and still not be ready.

Why did coding pull so far ahead? Gil asks whether it is looser criteria, a better training set, or something more like a research problem. Ng's answer is closer to markets than to architecture. The economic value of coding is clear, apparent, and massive, so resources poured in, and the smart people building the tools are themselves the users, which gives them product instinct for free. Gil presses: so this is capitalism at work plus domain knowledge, not a unique research obstacle? Ng does not flinch. "I think capitalism is great at solving fundamental research problems."

He is also watching models start to bootstrap themselves. Leading foundation-model labs have said publicly that they already use AI to write a lot of their code. What excites him more is agentic workflows that generate data for the next generation: Llama-style research in which an older model thinks for a long time to write hard puzzles, then you train the next model to solve them quickly. Multiple vectors, again. He rejects the phrase *vibe coding* in favor of AI-assisted coding. Vibe coding, he says, makes people think they can accept every suggested change and ride the mood. Sometimes that works. After an afternoon of real AI-assisted coding he is mentally exhausted. He calls it rapid engineering: still engineering, just much faster.

At AI Fund, which incubates startups, Ng is watching the core loop break its old proportions. Write software, test it with users, decide how to change it. Things that once took six engineers three months, a friend of his now throws together on a weekend. Coding speed is up and cost is down, so the bottleneck is product management: deciding what you actually want to build. If a prototype took three weeks, a week of user feedback was fine. If you can build the prototype in a day, waiting a week feels painful. His teams increasingly rely on gut, in a specific sense: they collect a lot of data that feeds a human mental model of the user, then they have to have deep customer empathy and make the call fast.

Tools that might speed up product work exist, and he is not dismissive. The Figma IPO is one proof that design tooling matters. There are systems that use AI to interview prospective users, and papers on flocks of AI agents simulating a target audience. It all feels promising and early. He does not think those tools are accelerating product managers nearly as much as coding tools are accelerating engineers, which means the bottleneck is getting worse. Gil mentions his partner Mike's view that computers can now interrogate humans at scale, and points to Listen Labs on the consumer-research side.

How rare is the product person who can live in that loop? Gil quotes a well-known public-company CEO who thinks there are at most a few hundred great product people in tech globally. Ng does not buy the census. There are, he says, way more than a few hundred great AI people too. What is scarce is the empathy: taking surveys, a handful of conversations, market reports, and behavior on neighboring apps, then getting out of your own head to form a model of an ideal customer you can decide for at speed. Early in his career he made a dumb mistake, training a batch of engineers to be product managers and mostly making excellent engineers feel bad for not being good PMs. The correlate he later trusted was very high human empathy.

Startups have a structural break here that Google does not. While you are early you can serve one user profile. Pick one human who is representative enough, or one hypothetical ideal customer, and build for that person, and you can go quite far. That is why the Cursor team can decide so instinctively, with less user interviewing: they *are* the typical user. Claude Code, despite huge usage, still was not folding large-scale user feedback into the training loop; direction was "what do we think the product should be at this stage." Building a startup, Ng says, feels more like tennis than calculus. You do not have time to work the proof. Most decisions are Jeff Bezos's two-way doors: if you are wrong, walk back through them a week later. Speed comes from obsessing, day and night, about the customer and the technology.

Ng has a habit of asking whether anything his teams are doing today they were also doing in 2022, and if so, whether it still makes sense. A lot of 2023 workflows already do not. Founders who are on top of the technology, tech-oriented product leaders, are much more likely to succeed than someone who is business-savvy but has no feel for where AI is going. Without that feel, he says, it is really difficult to think about strategy. Gil hears old Silicon Valley in the remark: Gates, Jobs and Wozniak, the early semiconductor and internet people, were highly technical. The industry forgot that for a while. Disruption puts the rare knowledge back in the people who can see the stack move.

Mobile is the contrast. Everyone roughly knows what a phone can do, what an app is, that GPS exists, so you do not need to be very technical to have a gut for "can I build a mobile app for that?" AI is not in that steady state. What a voice app can do, what foundation models will do next, what reasoning models unlock, is itself scarce knowledge. Technical fluency is a bigger differentiator now than it was in mobile. Gil notes that many of the largest mobile apps were still started by engineers: WhatsApp, Instagram, Instacart out of Amazon engineering, Travis Kalanick at Uber technically adjacent and early to GPS. You had to see the capability coming.

On founder temperament Ng says something he knows is not fashionable: working hard pretty clearly correlates with personal success. Not everyone is in a season where they can; the week his kids were born he did not work very hard, and that was fine. The factual pattern remains. Gil is blunter. He does not believe startups are for everyone. You are trying to do an unreasonable thing, create a lot of value very quickly, and unreasonable things take unreasonable hours. Ng quotes the line that the people who change the world are the ones crazy enough to think they can. He has seen two kinds of competitive founder who both work: people who desperately want their company to win, and people so obsessed with the customer winning that the company follows. Early Coursera, he says, he was the second type, not deeply informed about rivals, consumed by learners, and that obsession drove almost every decision.

The headcount-versus-AI story sits on a broader organizational bet. Some of the most productive teams Ng is on now are among the smallest: a few very good engineers, heavy AI enablement, low cost, together in person. Gil offers Open Evidence, which he says has reached about half of U.S. doctors and has made staying as small as possible an explicit company objective. Ng has heard that objective a lot, including through the 2010s. A bunch of companies under-hired, stayed proudly profitable, and grew less than they could have. Capital efficiency is wonderful until it is an excuse for going slower than the opportunity.

The failure mode, as Gil sketches it, is early-stage competition against other startups. Being far ahead feels great. Five, six, seven years in you hit incumbents, and the faster you captured the market and moved upmarket, the less time you gave them to notice. Slack versus Teams is one version of the story; Sketch, profitable and content in the Netherlands, then the Figma wave, is another. Minecraft, sold to Microsoft, is the counter-example people forget: a tiny team, a few billion dollars, massive use. You could always be lean. The question is leverage, distribution, and what you actually needed to spend on. Small teams are efficient with AI partly because small teams are efficient, period: you did not hire thirty extra people who get in the way. Many big tech companies, Gil suspects, could shrink by something like 70 percent and be more effective. AI also arbitrages high-value talent into markets that did not used to get it. Great engineers did not want to work in legal; Harvey changed that. Healthcare is seeing the same in-migration.

Software engineering, Ng thinks, is a harbinger. When Gil led Harvey's Series B he called customers in a profession famous for resisting new tools. The big firms furthest along all thought this was the future. What they worried about was organizational: if AI is everywhere and you hire ten associates instead of a hundred, where do future partners come from? Ng expects that reconstruction to replay industry by industry over a decade or two. He has a legal startup of his own, Calidas AI, coming out of AI Fund. The parts that will not automate soon, in investing as elsewhere, are where humans still hold an information advantage. A friend will make an offhand comment on a reference check that changes a decision, and that friend will talk to you, not to your model. Trust and relationships are the other durable piece. All the information in the world about why a role is a good opportunity is not the same as someone you trust saying you have to take it.

When new technology arrives, old excellence can become a liability. Gil sees engineering and go-to-market leaders still running 2022 playbooks inside companies whose ground is lava. Ng's analogy is web search: you would no longer hire someone, for many roles, who cannot search the web. He thinks we are well past the equivalent point for large models. At AI Fund, everyone knows how to code. Everyone has a GitHub account. When the assistant general counsel, the CFO, or the front-desk operator learns to code, they are not becoming software engineers. They get better at their actual jobs, because they can tell a computer more precisely what they want it to do.

That shows up in hiring. A research-engineer search pitted a strong ten-year full-stack résumé against a fresh college graduate. The experienced candidate had barely used AI tools; the graduate had. Ng hired the graduate and still thinks it was the right call. The other half of the story matters too. The best engineers he works with today are not new grads. They are people with ten, fifteen, or more years of experience who are also on top of AI tools, and they are in a different league. Experience plus AI compounds. Experience that pretends AI is optional does not. Gil has watched later-stage AI companies pick heads of product and design specifically for fluency with AI-assisted coding, because if you cannot mock the thing in the meeting you waste a week writing a requirements document.

For what to build next, he is skeptical of top-down "jobs at risk" lists, though he will glance at economists such as Erik Brynjolfsson for inspiration. The lesson inside AI Fund is an obsession with speed and with *concrete* ideas. "AI will transform healthcare" is true and useless. A subject-matter expert who points at one slice of healthcare operations and a vision model that might remove a step is something you can test: do customers want it, is it feasible, go. Deep research on companies and competitors looks ripe to automate; follow-on investment decisions for a portfolio of dozens of companies probably do not need full automation. LP reporting is a pile of paperwork that might be simplified. First-time technical founders, who now look like the core profile, still need help with the 20 percent they have not practiced, and venture studios get more reps at fundraising and customer-feedback tempo than any founder does. Over five years Ng expects people who embrace AI, in every job function and in personal tasks, to become far more capable than most people currently imagine. Two years ago, who would have guessed how productive software engineers would be today?