Why Howie Liu Sold Airtable to Build a Team of AI Agents

Liu says AI changes software economics. His new startup is betting that small businesses need coordinated agents, not another single assistant.

HLHowie LiuGast · Hyperagent
Kanal
Sources Podcast
Veröffentlicht
17. Nov. 2026
Video
58:22
Lesezeit
11 Min.
Original
Englisch
Auf YouTube öffnen
Beim Scrollen schrumpft das Video in die rechte untere Ecke und läuft weiter. Untertitel folgen Ihrer Lesesprache.
LesespracheIn 2 Sprachen verfügbar
Diesen Artikel gibt es noch nicht auf Deutsch, daher wird das Original (Englisch) angezeigt.
ZUSAMMENFASSUNG

In 30 Sekunden

Howie Liu sold Airtable while it was still growing because he saw a widening gap between the economics of its seat-priced software business and the opportunity he saw in AI. Hyperagent, incubated inside Airtable and spun out after the sale, is now being redesigned for teams of agents with separate roles and permissions. Liu hopes small businesses will use them for work they cannot afford to staff, but says the product must make setup simple while managing cost, mistakes and data access.

  1. 1Airtable was still growing when Liu chose to sellHe says the company was nearing $500 million in revenue but faced a different market from the one in which its per-seat SaaS business began.
  2. 2A cash-rich incumbent does not automatically make a better new productLiu says money and distribution cannot substitute for finding a product people want. Hyperagent is targeting small businesses for a faster learning cycle.
  3. 3Retained cash helped clear investor claims before common sharesLiu says Airtable had raised more than $1.4 billion but kept much of it, allowing the sale to return money to common shareholders after priority claims.
  4. 4Hyperagent is shifting from one assistant to multiple rolesLiu describes a planned product where agents have separate identities, permissions and ways to work with people and one another.
  5. 5Autonomy has to be balanced against cost and mistakesLiu says an agent can waste model usage, leak data or send an inappropriate message. A business product needs sensible defaults and boundaries.
  6. 6A landscaper could use agents from inquiry to job managementLiu sketches agents researching a property, helping prepare a quote and following field notes after a sale; he presents it as an example of work a small business could augment.

Das Wichtigste

  1. A high startup valuation does not ensure employee equity pays out: investor preferences can have priority over common shares.
  2. Liu says cash alone does not find product-market fit; the new product still needs careful design and iteration.
  3. A multi-agent business product needs identities, permissions and coordination as well as capable models.
  4. Liu sees small businesses as a faster feedback loop than large enterprise customers.
  5. Agent autonomy requires choices about spending, data access, errors and oversight.
VOLLTEXT

Den bearbeiteten Text nach Kapiteln lesen

Nur das Transkript
01

Why Liu sold a growing Airtable

Liu argues that Airtable was built for an earlier software era and that AI changes the economics of its next opportunity.

Howie Liu says Airtable emerged when powerful browser applications and product-led software were still new. By the time he sold it to Bending Spoons, the company had grown for more than a decade and the assumptions behind its business were changing.

When Airtable began in 2012, Liu recalls, products such as Dropbox and Evernote had shown that people could adopt software themselves. Slack had not yet launched. Browsers were also becoming capable of running applications that felt more like desktop software.

Companies including Airtable, Figma and Miro grew alongside those capabilities. Liu sees timing as part of Airtable's success: its product and business model made sense for a particular moment in software.

More than a decade later, Liu says, AI is changing distribution, pricing power and the economics of building a software business. He believes it has introduced a different technological foundation rather than simply another feature to add.

Airtable was approaching $500 million in revenue, still growing and generating cash, Liu says. But he thought its position was different from the beginning of the software-as-a-service boom, when the tailwinds that shaped it were stronger.

He contrasts per-seat software pricing, which typically sells productivity to employees, with AI systems that may automate work and incur variable model usage costs. A mature software company can add AI, he says, but its underlying economics and organization do not automatically change.

The choice, in Liu's account, was to continue operating Airtable as a SaaS business in the AI era or find a buyer. He came to respect Bending Spoons' approach to acquiring and holding businesses and believed it could run Airtable well.

Hyperagent had been incubated inside Airtable. The sale allowed it to spin out as a separate company, while giving Airtable shareholders a stake in the new opportunity, Liu says. That separated a mature software business from a bet built around AI from the outset.

02

Why Hyperagent became a separate company

Liu says capital and incumbent distribution are less decisive than the work of finding a new product and market.

Alex Heath asks why Liu did not retain Airtable and use its cash flow to fund Hyperagent. A profitable core business might appear to be an ideal source of capital for the new product.

Liu argues that capital is rarely the principal constraint when a team is trying to find a product people want. Well-funded startups can fail to build something compelling, and large incumbents can struggle to enter a new market despite their money and distribution.

For a product meant to be adopted without a sales process, he says, money can become a distraction or a substitute for the painstaking attention needed to get the experience right. The difficult product work still has to happen.

The businesses also have different economics, Liu says. AI usage can carry costs unlike the high-margin SaaS model, and the distribution strategy for a new agent product need not resemble the one used to sell established enterprise software.

Airtable had built a customer base that included large enterprises. Hyperagent, by contrast, is initially prioritizing small-business owner-operators, which Liu expects to let the team release changes and learn faster than it could through large enterprise sales cycles.

Heath asks why Bending Spoons was a suitable buyer for Airtable. Its portfolio includes established brands, and he notes that the company may be less familiar to listeners in the United States.

Liu says the buyer has a disciplined acquisition playbook: understand a company's economics, decide what it can pay and operate the business. He values a transparent counterparty that behaves consistently with its stated approach.

He contrasts that certainty with acquisition processes that consume time and expose company information without reaching a deal. Regulatory risk can also turn a proposed transaction into a lengthy distraction, as the failed Adobe-Figma deal showed.

03

What the sale meant for employee equity

The guests explain why investors’ priority claims matter to common shareholders, even when a company has a high headline valuation.

Leading a team through a sale is messy, Liu says. People make different choices about staying or joining a new venture, and leadership can require deciding on everyone's behalf without making every individual happy.

Airtable sold below its peak valuation. Liu says many companies reached unusually high valuations when interest rates were low; he thought waiting might produce a worse outcome if the market for SaaS or the competitive effects of AI deteriorated.

Airtable had raised over $1.4 billion across funding rounds, Liu says, but retained much of it and replenished cash through operations. That helped it repay investors' priority claims and return money to holders of common shares.

Heath stresses that employees need to understand how startup equity works. Investors may hold preferred shares with claims that are paid before common shares, so a sale or an impressive headline valuation does not guarantee employee equity has value.

Liu agrees that a rising company valuation is not the whole story. Different classes of equity can have different outcomes, and debt or investor preferences can complicate the value available to employees when a company is sold.

His rule of thumb is to preserve options for the company and its people. Raising a large round and spending it all creates a different situation from retaining a reserve or spending carefully: the preference still has to be cleared before common shareholders receive proceeds.

04

Coding agents brought Liu back to building

A holiday of experimenting with models led Liu from Airtable’s app-building platform to a product for building agents.

Heath turns to Hyperagent's origins. Liu says it began less as a grand plan than as a period of experimentation during the winter holidays, when he rediscovered how much he enjoyed building products himself.

Using coding agents and frontier models, Liu found he could make more than superficial prototypes. Drawing on his own engineering experience, he judged that the agents could help build working products with substantive underlying structure, faster than assembling a team for each experiment.

The experience reminded him of Airtable's early days. A person could try an idea without first formalizing a large project or hiring a team, though Liu notes that the model credits he used would not necessarily have been affordable to him as a student.

Liu distinguishes two approaches to product design. One begins with a customer's job and finds technology to serve it; the other begins with a new technical capability and explores what products it makes possible. His renewed interest in agents came from the second approach.

He had already seen earlier AI chat products create opportunities, but the newer coding agents felt like another step change. Rather than treating them only as a hobby, he began considering a product built around the capabilities he was using.

Airtable was software for building applications; Hyperagent grew from a similarly recursive idea: use agents to help people build agents. Liu followed the rise of OpenClaw and began learning by building the new product himself.

That work also made the differences from Airtable clearer to him. He says building an AI-native product involved different team sizes, resources, culture and business assumptions than those of the established SaaS business.

05

From one assistant to a team of agents

Liu outlines a planned redesign around multiple identities, permissions and collaboration among agents and people.

Heath describes Hyperagent's goal as multiple agents working together and asks how it fits among many competing agent products. Liu says he uses competing products extensively to understand what a good agent experience feels like.

In the older software market, Liu says, a capable product was difficult to copy quickly while its maker continued improving it. Agent products change faster, leaving incumbents less time to rely on existing distribution or yesterday's advantages.

Liu says Hyperagent 1.0 was valuable for learning agent mechanics and hearing from early users, but its successor is being redesigned around a different focus: small businesses employing teams of agents rather than one person using a single assistant.

Personal agents typically act alongside one person and use that person's accounts. Liu sees a different design problem when an owner wants several agents to work with several people, with separate roles, identities and permissions.

He compares the arrangement to hiring employees: a marketer should not send every message from the chief executive's account. Separate identities could make each agent's actions easier to audit and allow policies to differ by role.

The proposed product would let people communicate with several agents in group conversations and coordinate work among them. Liu presents this as the direction of Hyperagent 2.0, not as proof that every part of the design has already succeeded.

06

Simpler defaults and limits on autonomy

The product must balance easier setup with the cost and risks of agents acting independently.

Heath says the first Hyperagent offered extensive controls, down to spending, models and behavior, but could feel overwhelming. He asks why Liu is betting on multi-agent work when many people have not yet experienced a single personal agent.

Liu calls the personal-assistant market a knife fight in a phone booth. He believes large companies with access to models, distribution and substantial computing resources are all determined to compete there.

Multi-agent coordination is less settled, in his view. Designers have to decide how to divide work, share information and run tasks in parallel; there is no single obvious arrangement that fits every business.

He accepts the criticism that version 1.0 exposed too many settings. The new product needs sensible defaults that work out of the box, while still supporting a group of agents and people working together.

Greater autonomy creates tradeoffs. Liu says poorly configured agents could spend money on idle work, leak information, send awkward messages or take other harmful actions. The design must decide what they may do and how they coordinate.

He adds that differentiation is not enough if the basics fail: agents must also be capable, responsive and affordable. He sees the multi-agent category as a more promising place for Hyperagent to compete, without claiming a guaranteed outcome.

07

What small businesses might do with agents

Liu describes customer use cases and a landscaping workflow, from researching inquiries to tracking jobs.

Hyperagent's early audience includes small insurance brokers, landscapers, designers and agencies, Liu says. These owners often want another person to handle work rather than another narrowly specialized piece of software.

Liu says some customers use agents to augment staff and take on work they could not otherwise afford to hire for. He presents this as their reported experience, not as a measured result for every small business.

He gives a landscaping example: an agent could research an incoming inquiry, inspect satellite imagery and property photos, and help estimate a quote. It could also prepare an illustration and proposal for a human owner to use with the customer.

Once a job is won, agents could help track schedules and field notes, such as which plants or tools a crew needs next week. Liu says those operational details can determine whether a customer stays with a service provider.

The early users he describes are not all technical founders. They are willing to experiment, but may not be able to navigate deep configuration. That makes a self-service experience particularly important for the audience Hyperagent wants to reach.

Liu imagines small businesses gaining access to capable assistants at a cost below hiring an equivalent number of additional people. He expects the resulting gains to create a large market, although that outcome remains his projection.

08

A new startup amid a crowded AI market

Hyperagent raised funding, but Liu says the team must still prove its product. He and Heath examine the possibility of excess investment.

Hyperagent raised a round of many tens of millions of dollars shortly after the spinout, mostly from existing insiders, Liu says. He regards funding as less important than proving that the next product creates distinctive value for customers.

Version 1.0 has some traction, he says, but his team is treating the redesigned product as a new seed-stage effort. His previous experience at Airtable does not remove the need to find a product that works.

Heath asks whether the enthusiasm and valuations around AI resemble the low-interest-rate era that produced inflated SaaS valuations. As an investor, he sees many companies projecting rapid growth and wonders whether the optimism is too broad.

Liu prefers to reason from AI's underlying capabilities rather than rely on a direct analogy. He expects the technology to have unusually broad economic effects, while acknowledging that forecasts about growth and valuations are difficult to make precisely.

He concedes that capital can flow into companies that will not become big winners. It remains uncertain how much value frontier model providers will capture and how much will be available to younger application companies.

Liu says earlier SaaS valuations rested partly on expectations of high margins and cheap distribution after the software was built. Some of those assumptions may no longer hold in the same way for products that consume AI resources as they operate.

09

The old SaaS rules are no guarantee

Liu says software founders should reconsider whether familiar business assumptions still hold.

Asked what established SaaS founders should do if they cannot sell and start again, Liu says not every company is guaranteed to thrive through a technological transition. Leaders should reconsider whether the old rules still support their business and their employees' equity.

He argues that being skilled at the previous business model does not prove it will retain its value in a changed market. In his view, founders should compare the prospects of continuing with the alternatives rather than assuming that inaction is risk-free.

10

How Liu uses agents at Hyperagent

He describes agents for marketing, competitive research and coding, and his hopes for around-the-clock work.

Heath asks how Liu uses Hyperagent himself. Liu says he turns to agents for strategy, product and marketing questions, and describes using one to turn an advertising concept into a video rather than merely discussing the idea.

He also describes an agent that follows competitors, studies their releases and, with access to the codebase and development discussions, can propose a feature and produce a pull request for the team to inspect.

Liu contrasts that with a conventional sequence of documents, meetings, prototypes and decisions before a feature ships. He values being able to explore a concrete implementation earlier, though the resulting work still needs to be evaluated.

Agents can work overnight and accept requests at any hour, he says. They do make mistakes, but he values their patience in repeatedly checking a current feature against competing products and carrying out a detailed comparison.

Liu says his team is learning these practices by using its own product. He predicts that businesses in fields such as landscaping and insurance could gain much more speed and efficiency from well-deployed agents; his estimate of the advantage is a forecast, not a measured result.

Mehr aus diesem Kanal

1:08:18EN
Sources Podcast· 3. Nov. 2026Podcast · 2 StimmenAlex Davis on Concentrated Investing, AI Compute, and Private-Market RiskGast: Alex Davis · Moderation: Alex HeathThe Disruptive investor explains his new fund, his approach to late-stage companies and his concerns about speculative infrastructure and private-market vehicles.Venture CapitalInvestingAI Strategy+1
1:24:20EN
Sources Podcast· 28. Okt. 2026Podcast · 2 StimmenTobi Lütke on AI Agents, Entrepreneurship and Shopify’s FutureGast: Tobi Lütke · Moderation: Alex HeathShopify’s CEO talks with Alex Heath about agent-led shopping, company culture, monetization and the AI systems he runs at home.Agentic commerceEntrepreneurshipAI Adoption+2
51:38EN
Sources Podcast· 16. Juli 2026Podcast · 2 StimmenTolan Founder Quinten Farmer on AI Guides, Real Relationships and Building a Calm CompanyGast: Quinten Farmer · Moderation: Ellis HamburgerFarmer explains the alien character’s design, user-reported outcomes and premium pricing—and why he wants it to challenge users as well as listen.AI agentsProduct DesignCompany Building+1
Folien1:02:09EN
Sources Podcast· 9. Juli 2026Podcast · 3 StimmenJon Steinback on Marker, AI and the Thinking Behind WritingGast: Jon Steinback · Moderation: Alex Heath, Ellis HamburgerThe Marker founder argues for an editor that asks questions instead of writing for the user; the hosts examine where AI fits in their own work.Product DesignCreativityAI Product Management+1
49:05EN
Sources Podcast· 24. Juni 2026Podcast · 2 StimmenJason Yuan on Hivemind, Social AI and the Risks of Sharing Between FriendsGast: Jason Yuan · Moderation: Ellis HamburgerThe founder describes an AI that asks people in a network for answers, creates opportunities for connection and must navigate imperfect privacy safeguards.AI agentsPrivacy & data governanceProduct Design+1