HyperAgent: Airtable CEO Howie Liu Just Gave Away the Agent Playbook

Howie Liu doesn't need your money.

Airtable does half a billion in revenue, sits on a billion dollars of cash, and will throw off $100 million in free cash flow this year.

So when he goes on Greg Isenberg's podcast and gives away $1 million in credits for his new product, HyperAgent, it's worth asking why.

The answer, in his words: the agent opportunity isn't the trillion dollars Sequoia puts on it. It's "the whole GDP of all white collar labor." Tens of trillions. And he thinks solo founders will move on it faster than any 50,000-person company can.

I watched the full episode so you don't have to (though you probably should). Breakdown below.

The market read from a founder who's already won

Greg opens with a Sequoia chart: where are AI agents deployed today? Software engineering sits at almost 50%. Back office at 9%. Marketing at 4%. Sales at 4.3%.

Howie's take surprised me. He thinks even the 50% is an overestimate.

Because most engineers "using AI" are still in what he calls Gen 1 mode: autocomplete in an IDE, the stuff we figured out three years ago. The actual frontier looks different. Howie builds HyperAgent with 30 Claude Code instances running in parallel, each hooked to a browser, fully autonomous, opening PRs and getting other agents to review them.

Almost nobody works like that yet. Which means every single category on that chart, including the "mature" one, is still early.

Two more numbers he dropped that frame the opportunity:

→ OpenAI and Anthropic went from roughly zero to a combined $80B+ in revenue in a few years. No software category has ever done that.

→ A recent board memo Howie sent cost him about $150 in tokens to research and draft with HyperAgent. His best investors called it the best memo he'd ever written. It took a tenth of the usual time.

That $150 detail matters more than the trillion-dollar talk. People anchor AI pricing to Netflix subscriptions and flinch at token bills. Wrong frame. Anchor it to what a human would cost to do the same work, and $150 for your best-ever board memo is basically free.

What HyperAgent actually is

Howie's pitch: if OpenClaw and the other agent tools are Linux, HyperAgent is the Mac.

Cloud-native, so no Mac mini humming in your closet. Secure by default. And carrying the same UX obsession that took Airtable from "databases are technical" to a product normal people use.

The demo made it concrete. Greg pitched one of his startup ideas: hyper-local market reports for real estate agents, built from public data. HyperAgent went off and:

• Researched the market and sized it (a "couple billion TAM": big enough to matter, small enough that incumbents ignore it)
• Pulled real Reddit threads of real estate agents asking for exactly this product
• Ran a competitive analysis
• Then built a working V1 of the app

Howie's framing on that last step stuck with me: "App building is the feature now. It's a commoditized feature." Any frontier agent on a frontier model builds a clean app. The difference is doing the founder work first, the research and validation, and then building the app informed by it.

Howie positions HyperAgent as your co-founder, not your developer.

Skills: the most important primitive in agents right now

If you take one concept from the episode, take this one.

Howie calls skills "the most important primitive in the frontier agents world." His analogy: the models are already Einstein-smart in a general sense. Einstein still can't do your real estate job without a briefing. A skill is that briefing, written down, reusable, and improvable.

Live on the call, he built a skill that drafts tweets in Greg's voice. It researched Greg's actual posts and distilled rules like "Greg's voice is a smart friend at dinner saying the quiet part out loud" and "hook in the first seven words."

The first drafts? Fine. A bit stiff. Maybe 50% of the way there.

And that's the point where most people quit. Howie was blunt about it: people one-shot something, get a mediocre result, and walk away concluding agents are overhyped. The capability was never the problem. The coaching was.

You give feedback ("these feel too formal, make them friend-to-friend"), the skill updates, and every future run starts from the improved version. It compounds like an employee who never forgets a correction.

Rubrics: managing agents like a CEO manages people

This is the part built for anyone trying to run a fleet, not a toy.

A rubric in HyperAgent is an eval: you define what good looks like for a task, and a separate model scores every output against it automatically. LLM as judge, running in the background.

Why bother? Because if you're personally reviewing every output from every agent, you're not running agents. You're babysitting them. Greg made the comparison to Manus: the judge of the output is you, a human, and you don't scale.

Rubrics also cut costs in a way I hadn't considered. Watch the score trend line, then try dropping the model from Opus to Sonnet. If quality holds, you just cut that agent's cost by 5x with data to prove it was safe.

It's management 101 applied to software. You don't check every employee's every task. You build checks that scale. Allie K. Miller runs her 34-agent operation on the same logic, with watchdog agents instead of dashboards (I broke her system down in this story).

The rest of the toolkit

Quick hits from the demo that founders will care about:

• One-click Slack deployment. Any agent becomes an always-on coworker in your channels, listening and chiming in when relevant.
• Connectors plus custom API skills. OAuth into Gmail, Notion, Slack, Granola out of the box. No connector? Howie had it read the Twilio docs live and build its own integration skill, then ask for credentials safely.
• Memory defrag. Shipped the day before recording: it clusters related memories across agents by keywords and embeddings so your fleet's context stays clean as it grows.
• Onboarding that researches you. Connect your accounts and it reads your last weeks of email, Slack, and meetings, then suggests agents worth building. The VC example: an agent that auto-researches every inbound pitch and threads a private summary into the email.
• Live mode (coming). Heartbeat-style always-on agents that push ideas to you over Telegram, email, or Slack.

On the money side: first 1,000 people from the episode get $1,000 in HyperAgent credits. Real credits, usable on frontier models including Opus, subsidized by Airtable's balance sheet. Howie is explicit that he's buying adoption the way Airtable did with PLG, because he wants HyperAgent to be "the iPhone" of agents.

The 30-minutes-a-day arbitrage

The best exchange of the episode had nothing to do with the product.

Howie asked Greg: when does this stuff start paying for people? Greg's answer maps to what I hear from founders on the podcast every week.

Your first internet dollar rewires your brain. At $10K a month, you quit your job.

And the way you get there with agents isn't a heroic weekend. It's 30 minutes a day, on the calendar, for 30 to 90 days straight. Greg's claim: that alone puts you in the top 1% of agent builders, because 99% of people try it sporadically, get a mediocre first output, and quit.

Howie backed it with a parable from a Benchmark partner memo. Two knife salesmen in 2003. One spends nights tinkering with this new Google AdWords thing on the side. The other stops selling knives entirely, eats his savings for two months, and figures out ecommerce properly. Five years later one of them owns a massive online business and the other is still knocking on doors in a shrinking market.

The uncomfortable part of that story is the two months of zero revenue. The learning period costs something. It's still the cheapest thing you'll ever buy.

That's the actual pitch of this episode, and it applies whatever tool you pick. The skill of running agents (writing skills, setting evals, managing a fleet) is what compounds. I made the same argument in why managing AI agents is the most valuable skill of 2026, and Howie, sitting on a billion dollars, just agreed.

FAQ

What is HyperAgent?

HyperAgent is an AI agent builder from Airtable, led by co-founder and CEO Howie Liu. It runs agents on cloud computers with full coding ability, built-in connectors, skills you can train, rubric-based evals, and one-click deployment into Slack. Howie describes it as the Mac to OpenClaw's Linux.

How is HyperAgent different from OpenClaw, Manus, or Perplexity Computer?

Versus OpenClaw: turnkey, cloud-native, and secure by default, with no self-hosting. Versus Manus and Perplexity Computer: more powerful tools out of the box, plus a fleet layer (command center, rubrics, Slack deployment) designed for running many agents, not just chatting with one.

How much does it cost to run agents like this?

Anchor on human-equivalent cost, not subscription pricing. Howie's board memo cost about $150 in tokens and replaced days of work. Rubrics help you cut costs over time: if a cheaper model holds your quality score, you drop from Opus to Sonnet and save around 5x on that agent.

How long until agents actually pay off?

Greg's rule from the episode: 30 minutes a day, every day, for 30 to 90 days. That consistency puts you ahead of the 99% who try agents sporadically and quit after one mediocre output. First milestone is your first internet dollar. The quit-your-job milestone is around $10K a month.