Making $$ with AI Agents
TL;DR
Howie Liu thinks the AI agent market is far bigger than Sequoia’s $1T estimate — he argues the real upside is closer to “the whole GDP of white-collar labor,” i.e. many tens of trillions, because frontier models are now smart enough to do expert-level work across functions, not just coding.
The big shift is from copilot to autopilot — Liu says software engineering’s reported ~50% agent penetration is still an underestimate because frontier teams have already moved past GitHub Copilot-style assistive AI into fully autonomous workflows with “30 different cloud code instances running in parallel.”
Token costs only look expensive if you compare them to SaaS, not humans — Liu’s example was a board memo partly crafted by HyperAgent that his investors called his best ever, and even if it cost $150 in tokens, it saved roughly 10x in executive time.
HyperAgent’s pitch is ‘the Mac version’ of agent builders — compared with OpenClaw’s Linux-like power-user vibe, Liu frames HyperAgent as cloud-native, visual, secure, and designed to make agents feel intuitive enough for non-technical builders while still scaling to fleets of agents.
The product isn’t just an app builder; it’s trying to act like a founder — in the demo, HyperAgent researched Greg Isenberg’s real-estate startup idea, validated demand with Reddit and market data, wrote a business brief, then built a clean V1 app informed by that research.
The real arbitrage is skill-building, not one-shot prompting — both Greg and Howie hammer home that most people quit too early, while the winners will be the ones who use agents daily for 30, 60, or 90 days, refine skills and rubrics, and gradually build agent-first businesses with tiny teams.
The Breakdown
Sequoia’s charts, and why Howie thinks even the bullish numbers are too small
Greg opens in full “agent psychosis” mode, tossing up Sequoia charts showing software engineering at nearly 50% of AI agent deployment while most other functions are still single digits. Howie Liu’s take is basically: this chart measures how early we still are, not how far we’ve come — if frontier agents were actually deployed seriously, every category should be much closer to 100%.
From AI autocomplete to fully autonomous work
Howie draws a sharp line between old-school AI augmentation and the new autonomous workflow. He says top teams are no longer just using Copilot in an IDE; they’re running parallel coding agents in the cloud, hooked to browsers, reviewing PRs, and shipping work that feels like it came from a real engineer — echoing Andrej Karpathy’s shift from mostly human-written code to mostly AI-written code.
Why the opportunity is bigger than $1 trillion
When Greg cites Sequoia’s “$1 trillion up for grabs” figure, Howie immediately zooms out: if agents can do white-collar work, the TAM is not $1T, it’s many tens of trillions. His point is that model intelligence has crossed a threshold in the last 4-5 months — he calls out Opus as a breakthrough — and now the real bottleneck is adoption, not capability.
Stop whining about tokens and compare them to labor
One of the most useful reframes in the episode is around cost. Howie says people are still mentally anchoring AI pricing to cheap subscription software, when they should be comparing it to what a human would charge or how much executive time it saves; his own example is a board memo largely researched and shaped by HyperAgent that got rave reviews from investors.
Enterprises are spending because doing nothing is riskier
Greg then moves into enterprise adoption and agent command centers, and Howie says the growth curve doesn’t surprise him at all. He thinks one of the biggest cash grabs in business history is either product-led AI tools that spread because they work, or top-down “Palantir-style” enterprise pitches where CEOs will gladly spend $100M+ because the game theory is brutal: maybe you waste the money, but if you ignore AI, you definitely get fired.
Why agents are starting to look like org charts
The conversation gets philosophical when Howie explains why fleets of agents increasingly resemble human job roles. His analogy is robots converging on humanoid form factors because our world is built for humans; similarly, companies are built around specialized roles, so agents naturally map to content marketer, researcher, support rep, and so on instead of becoming one omnipotent blob.
The HyperAgent demo: research, build, deploy
Then comes the product demo. Howie describes HyperAgent as the “Macintosh” to OpenClaw’s Linux — same underlying frontier-model power, but with UX obsession, visual workflows, and cloud-native setup — and shows it taking one of Greg’s startup ideas, researching the market, validating pain points on Reddit, generating a business case, and then building a polished V1 app.
Skills, rubrics, and the real path to outperforming 99%
The most practical section is about “skills,” which Howie calls the key primitive in agent systems: the model is already smart, but skills are the playbooks that make it useful in a domain. He demos an agent that studies Greg’s X posts to create a reusable “Greg Isenberg contrarian AI takes” skill, then explains how rubrics let you score output quality over time — and both of them land on the same takeaway: the edge goes to people who stick with this daily, through the awkward stage, until the system compounds.