Airtable founder & CEO Howie Liu traces the progression of AI product form factors — from completions to chatbots to agents to always-on "claws" to agents that orchestrate other agents — and argues the next leap is hiring employable agents. Demoing HyperAgent through a real landscaping business, he shows agents that recurse, learn from feedback, and hand off work while humans stay in the loop to unblock high-impact actions. His close: humans become orchestrators of agent fleets, and adopting this is "table stakes for survival" for every company.
Liu opens by positioning himself as a "product form factor thinker at heart" rather than a technical deep-diver, and frames the talk as a high-level view of where agents are heading as a product-building matter. He lays out a spectrum that tracks rising model intelligence: completions (a model that just continues emitting tokens until a stop token), then chatbots (unlocked by instruct-tuning and the conversational back-and-forth that InstructGPT/ChatGPT enabled), then agents.
For "agent," he explicitly borrows Anthropic's definition — a model that recurses upon itself, with an open-ended set of tools it can call, reasoning steps, and decisions it can make — as distinct from a prescripted linear workflow. He notes the term has become badly overloaded over the conference's few days.
Liu distinguishes reactive agents — which respond to user input, perform a task for perhaps 20 minutes to an hour, and then stop — from what he calls claws, a term he attributes to the popularity of openclaw. Claws perform work for much longer and use a heartbeat mechanism (now also in products like HyperAgent) to wake up on their own and exhibit always-on behavior.
Under the hood, each wake-up drives many turns: the agent loops on itself, each tool-call response reinvokes the LLM until a stopping point, and the heartbeat restarts it toward its goal. He contrasts this with early experiments like BabyAGI and AutoGPT, which were fun to watch but inevitably drifted into rabbit holes — whereas today's models are smart enough to stay coherent on longer, more open-ended tasks and to orchestrate other agents (agent-to-agent) for advanced, longer-running jobs.
Liu grounds the abstraction in a real HyperAgent customer — a landscaping business that runs virtually every part of its operation with agents, from finding clients and building quotes and pitches to managing the physical work. He notes the prototypical HyperAgent users aren't only AI-forward software companies but often traditionally offline businesses, and credits the ingenuity of agent builders across every sector.
In the walkthrough, an inbound inquiry (a photo of a messy backyard and a request for a quote) is routed to a triager agent that assesses scope and lead quality, then hands off to a surveyor agent. Because every agent thread gets a fully capable sandbox VM, the surveyor can write code and use tools like ffmpeg to clip screenshots from a submitted video, analyze the imagery, and produce a high-touch proposal and interactive pitch deck — a level of bespoke output that Liu says would previously only have been economical for a multi-million-dollar client.
The demo pauses at a human checkpoint: the agent ("Sage") returns to the business owner with a surveyed lead, a pitch deck, and a realistic quote for approval before sending. Liu cites a recent Sarah Guo post arguing that as models get smarter, humans who still own policies and decisions remain the ultimate gate for high-impact actions — like issuing a binding quote — intervening like a manager receiving upward reports from a team.
He also argues agents must not be static. Echoing Gary's earlier talk about a sticky, learning context layer, Liu says agents should accumulate memories and skill updates and learn in real time from user feedback delivered via Slack or email, rather than waiting for the next fine-tuning run — remembering corrections like a fluid, evolving human collaborator.
Liu closes by drawing an analogy to coding's evolution — from single-threaded human coding, to completion-style autocomplete (first-generation GitHub Copilot), to chat agents (Cursor Composer), to today's frontier developers overseeing a whole fleet of agents. Going to sleep without setting agents to work overnight, he says, now feels like taking a huge loss.
The needed shift is a UX change: an orchestration control plane that shows, at a glance, what every agent is working on, what it's blocked on, and how work is handed off — letting the human zoom out to a "Sim City" macro view instead of micromanaging each task. He frames this leap into humans-as-orchestrators as not just an opportunity but "table stakes for survival" in every industry, invoking Jensen Huang's line that it's not AI but someone using AI who takes your job. He ends with a HyperAgent offer of $1,000 in inference credit.
really I'm a product form factor thinker at heart
my definition of agent really is the same as the anthropic definition, which is distinct from workflows
going to sleep without setting off your agents to perform useful work overnight feels like you're you're taking this huge loss
it's not going to be AI taking your job. It's going to be, you know, somebody using AI who takes your job or takes your business
| Time | Topic |
|---|---|
| 00:00 | Intro; the spectrum of AI form factors, completions to chatbots |
| 03:08 | Defining an agent (Anthropic's definition) vs. workflows |
| 04:08 | Always-on "claws" and the heartbeat mechanism |
| 06:43 | Airtable's horizontal philosophy; the landscaping use case |
| 09:16 | Triager and surveyor agents build a high-touch proposal |
| 12:19 | Human unblocking, real-time learning, and memory |
| 15:26 | Orchestration control plane; table stakes for survival |