Howie Liu co-founded Airtable and founded HyperAgent. His work spans two ways of making business software accessible: letting people build applications around their own data, and giving them agents that can carry out work. His product decisions connect that ambition to concrete requirements: linked records for modeling a business, inspectable logic for AI-built applications, and human approval when agents need help proceeding.
Building software people could entrust with their work
Howie’s first startup, Etacts, helped people maintain personal and professional relationships. Founded with Evan Beard and part of Y Combinator’s Winter 2010 batch, it surfaced neglected contacts, offered reminders to reconnect, and supplied context from earlier correspondence. Salesforce acquired the company roughly a year after it began, and Howie subsequently worked there in product. His Stanford interview traces that progression to Airtable, which he founded with Andrew Ofstad and Emmett Nicholas.
Airtable developed at a different pace. The team wanted people to build applications around their own work rather than fit every business process into a predetermined product. They spent approximately two and a half years developing it before its public launch in 2015.
An early commitment to a relational database shaped that effort. Linked records let users represent relationships between customers, projects, inventory, and other parts of a business; a familiar visual interface made that structure approachable. Howie regarded the underlying data model as a difficult decision to reverse later, worth resolving before chasing growth. His account of Airtable’s early development also describes the hands-on design work: he and Andrew carried large external monitors between their living rooms as they worked on the product.
That slower development reflected what customers would have to risk by adopting Airtable. They would entrust it with valuable data and important workflows, often replacing mature tools. Howie’s distinction between two kinds of minimum viable products explains the tradeoff: a lightweight experiment can test some ideas, while a business-critical replacement needs enough completeness and reliability to make switching reasonable.
Growth brought a consequential change in commercial focus. In December 2022, Howie announced a shift toward large enterprise customers, accompanied by 254 job cuts. He and Andrew accepted responsibility for the decisions leading to the restructuring. Airtable retained its broad application-building capabilities while narrowing the customers around whom it organized the business.
Inspecting what AI builds. A convincing generated interface can conceal faulty logic or insecure data handling. Airtable’s answer was Omni, an agent that assembled established platform components. Users could inspect the resulting data, logic, and layouts, making conversational creation compatible with understanding how an application worked.
Turning an answer into an ongoing operation. Researching thousands of companies consistently, then processing new arrivals automatically, requires stored data, triggers, and coordinated steps. Howie’s approach connected model-driven work to applications and human workflows so that an agent’s output could become part of a repeatable business process.
Agents that keep working between assignments
HyperAgent extends this work into agents with their own computing environments, browser and shell access, and assignments arriving through channels such as Slack, Telegram, schedules, and webhooks. The product emphasizes visible execution, persistent memories and skills, and deliverables that can be updated as circumstances change. Those deliverables include websites, documents, dashboards, slides, and video.
Howie’s argument for “employable agents” focuses on what agents need to participate in continuing work. He describes a progression from completions and chatbots to agents that remain active between requests and coordinate other agents. A landscaping-business example makes that distinction concrete: a triage agent and a surveyor agent divide the work, with human approval helping agents proceed when needed. His emphasis on learning from feedback adds another requirement: interactions should improve subsequent work rather than disappear when an assignment ends.
Howie Liu explains the product mechanics behind agents that keep working, hand off tasks and ask for approval. Three startup pitches then test where those capabilities create a business—and where domain knowledge, shared context and audience limits still matter.
Tool-result turns keep an agent working within a run; a heartbeat gives it another run after it stops. Neither mechanism alone ensures useful progress.
▶ Watch ↗