How Agentic AI is Transforming The Startup Landscape with Andrew Ng

Summary

• AI progress is broadening beyond scale, where Andrew Ng sees only “a little bit more juice out of the scalability lemon” and rising difficulty. Agentic workflows, multimodal systems, concrete applications, and possible wild cards such as diffusion models for text now matter alongside larger models. The business progress is real, but “the marketing hype has grown insanely fast.”
• Ng calls skilled implementation the single biggest barrier to agentic applications, even as technical components still need work. Computer use, guardrails, and evals remain imperfect, but the decisive capability is running “a systematic error analysis process with evals” against proprietary business context. For the next year or two, human engineers and product managers remain essential because the required knowledge often lives in employees’ heads, not internet training data.
• Coding agents are the clearest proof that highly autonomous agents can create substantial economic value today. Ng places coding beside ChatGPT-style question answering as AI’s two obvious value pools and calls Claude Code his current favorite because it can plan, build a checklist, and execute multiple steps. He rejects “vibe coding” as misleading: serious AI-assisted development is a “deeply intellectual exercise” better described as “rapid engineering.”
• Rapid engineering changes startup economics by moving the bottleneck from writing software to deciding what deserves to be built. Work that once required six engineers for three months can sometimes be completed by one person over a weekend; when a prototype takes one day but user feedback takes a week, product judgment becomes painfully scarce. Simulated users and AI-led interviews look promising, but product tools are not accelerating product managers nearly as much as coding tools accelerate engineers.
• Ng says technically fluent product leaders are now much more likely to succeed than business-savvy founders who lack a feel for a rapidly changing capability surface. Teams should inspect anything still done as it was in 2022 because much of it may no longer work in 2025. Customer empathy, speed, conviction, and hard work matter because startup decisions resemble “playing tennis” more than solving calculus problems: founders need enough accumulated instinct to act immediately through many reversible “two-way doors.”
• In one hiring example, AI fluency outweighed tenure, while experience plus AI fluency creates an even more formidable talent tier. Ng hired an AI-native college graduate over a full-stack engineer with 10 years’ experience who barely used AI tools, yet says the best engineers are veterans with 10–15-plus years who also master the new tooling—“completely in a class of their own.” At AI Fund, even legal, finance, and front-desk staff learn to code so they can specify work precisely to computers.
• Smaller AI-enabled teams can outperform, but optimizing for head count or profitability can become a strategic trap. Ng increasingly asks whether a task needs budget to “hire AI,” while Elad Gil warns that underhiring can give incumbents time to win through distribution, citing Slack versus Teams and Sketch before Figma. The calibration depends on market structure: in winner-take-all categories, speed to capture the market can matter more than keeping teams lean.
• The most durable human advantages are proprietary context, relationships, and judgment—while the upside belongs disproportionately to people who embrace AI. Competitive research and LP paperwork look automatable, but founder assessment still draws on offhand reference comments, in-person leadership signals, and trust that an AI cannot yet access. Ng’s five-year call is that adopters across job functions will become “far greater” in individual capability than most people currently imagine.

Deep dive

1. Scale is no longer AI’s only credible vector of progress

• Ng’s starting point is deliberately plural: there is “a little bit more juice out of the scalability lemon to be squeezed,” but extracting it is becoming “really, really difficult.” Scale dominates public imagination partly because a handful of companies with excellent PR made it the industry’s central narrative.

• Ng introduced “agentic AI” to end an unproductive binary argument over whether a system qualified as an agent. Autonomy exists on a spectrum—from prompting that merely affects an LLM’s output to systems that plan, reason through multiple steps, and execute independently—so his proposal was to call the whole spectrum agentic and “spend the time actually building this.”

• What he did not anticipate was marketers turning the term into “a sticker to stick on everything in sight.” His calibration matters: real business adoption is growing rapidly, but not nearly as rapidly as the hype.

• Other vectors include multimodal-model design, application engineering, and new technical approaches. Ng flags diffusion models generating text as an intriguing wild card and points to an older Llama model generating puzzles for the next generation to solve quickly—AI helping create training data for its successor.

2. Agent deployment is constrained by eval discipline and private context

• Computer use “kind of works, often doesn’t work,” while guardrails and evaluation remain significant technical problems. Yet Ng’s “single biggest barrier” is talent: strong teams systematically identify what works, what fails, and what to improve; inexperienced teams try changes more randomly and take much longer to converge.

• His representative workflow is “next gen robotic process automation”: receive a document, convert it to text, search the web for compliance issues, check pricing against a database, route it for verification, and save the result. When it fails, the consequential question might be whether an invoice date is wrong—or whether the system keeps bothering the CEO for approvals.

• Those distinctions depend on proprietary context, not general internet knowledge or a neatly extractable manual. Human product managers and engineers must decide which errors matter, whom the workflow may interrupt, and what the business will tolerate; Ng expects “a lot of work for human engineers” over at least the next year or two.

• The clearest working example is coding. Ng identifies ChatGPT-style question answering and coding agents as two massive economic-value pools, with Claude Code his current favorite for autonomous planning and execution; shopping and browser-control agents remain “really nice demos” rather than production-ready systems. When Sarah asks whether coding’s lead is capitalism plus domain knowledge, Ng replies: “Capitalism is great at solving fundamental research problems.”

3. Faster code makes product judgment the scarce startup resource

• Ng resists “vibe coding” because it implies accepting whatever changes the model suggests. His own AI-assisted sessions are mentally exhausting, “deeply intellectual” engineering: AI makes serious systems much faster to build, but it does not remove the need to understand and direct them.

• At AI Fund, work that might once have required six engineers for three months can now sometimes be built by Ng or a friend over a weekend. That compresses the startup loop unevenly: coding becomes faster and cheaper, while deciding what users want does not automatically improve.

• The resulting bottleneck is product management. A one-week feedback cycle was tolerable when a prototype took three weeks; it is painful when the prototype takes a day. Ng’s teams therefore collect data to build a mental model of the customer, then increasingly rely on gut and “deep customer empathy” to make decisions at software speed.

• AI-led user interviews and “a flock of AI agents” simulating a market look promising but early. Sarah’s broader formulation is that “computers can now interrogate humans at scale,” yet Ng does not believe these tools accelerate product managers nearly as much as coding agents accelerate engineers.

4. AI-native founders pair technical instinct with customer obsession

• Ng asks whether anything a company still does as it did in 2022 makes sense in 2025. In a fast-moving capability environment, he says technically oriented product leaders are much more likely to succeed than business-savvy founders who lack a good feel for what AI can and cannot do.

• Mobile eventually became legible enough that nearly everyone understood apps, GPS, and the device’s constraints; AI remains rare knowledge because its frontier keeps moving. Elad connects this to earlier technical founders, while Ng emphasizes the enabling insight: Travis Kalanick had to recognize early that GPS made a new service possible.

• On founder temperament, Ng says “working hard probably correlates to your personal success,” while acknowledging that people pass through periods when they cannot. Sarah sharpens the claim: startups are not for everyone because creating substantial value unusually quickly is “a very unreasonable thing.” Elad adds competitiveness; Ng distinguishes founders who obsess over beating rivals from those who obsess over making customers win, recalling Coursera’s early learner focus.

• Startup judgment, in Ng’s metaphor, is “more like playing tennis than solving calculus problems”: there is rarely time for complete analysis. Many choices are Bezos-style “two-way doors,” so founders can decide, reverse course a week later if needed, and earn speed by obsessing over the customer and technology.

5. AI literacy reshapes hiring, roles, and the optimal team size

• Everyone at AI Fund has a GitHub account and knows how to code, including legal, finance, and front-desk staff. They are not software engineers; learning “the language of computers” simply lets them tell machines more precisely what work to perform, making them better at their primary functions.

• Ng calls the environment “the floor is lava”: leaders who built careers around older ways of operating may be less effective as the pace of capability and competition changes. For many roles, he says, inability to use LLMs effectively now makes someone much less effective than a person who can.

• Ng’s hiring specimen is stark: an AI-native new graduate seemed likely to be more productive than a full-stack engineer with 10 years’ experience who had barely used AI tools, so he hired the graduate. The strongest engineers he works with, however, are people with 10, 15, or more years of experience who also master AI; he expects software engineering to foreshadow other professions.

• Harvey’s law-firm customers were already asking what happens when ubiquitous AI reduces hiring from 100 associates to 10: without the old pyramid, where do future partners come from? Ng wonders whether a “really small, really skilled team” with extensive AI support can outperform a larger outsourced workforce, particularly once lower coordination costs are included, but calls the destination uncertain.

• Elad’s pushback on lean-at-all-costs is the crucial caveat: startups can underhire, admire their profitability, and surrender the window before incumbents deploy distribution. He points to Slack versus Teams and Sketch before Figma; Ng agrees teams can be smaller, but winner-take-all markets may demand speed rather than a fixed head-count target.

• Ng’s own heuristic is revealing: he rejected a request for more human headcount but approved a request to “hire AI.” Recognizing when to add AI rather than more people is itself an important operating instinct.

6. Concrete workflows beat sweeping sector theses

• Ng consults economists studying which jobs face disruption, but finds top-down declarations such as “AI will transform healthcare” operationally useless. His broader lesson is that “AI will target rich environments.” AI Fund wants a subject-matter expert to identify a specific healthcare operation and mechanism; once an idea is concrete, the team can quickly test customer demand and technical feasibility.

• Within investing, deep company and competitive research look ripe for automation, and Ng already uses deep-research tools for cursory market work. LP reporting also contains “massive amounts of paperwork.” Follow-on decisions are less obvious automation targets: the firm makes them infrequently, has already reviewed the companies, and Ng says full automation is probably unnecessary.

• Ng identifies a relationship advantage in founder assessment: a human can notice leadership behavior in a meeting or catch an offhand reference-check comment that never reaches a model. Trusted advice like “Sally, you gotta do this. It’s gonna work” is not interchangeable with exhaustive information about an opportunity.

• First-time technical founders can close gaps through experienced peers, complementary hires, and investors or studios with more repetitions in feedback, fundraising, hiring, and speed. Ng also favors learning by doing: founders will screw things up, but that is acceptable when mistakes are not existential.

• Ng’s final call broadens beyond startups: people who embrace AI across work and personal tasks will become “so much more powerful and so much more capable” than most currently expect.