Hiring Employable Agents: Howie Liu's Agent Stack

Hiring employable agents describes a working pattern, not a product tier. An always-on agent wakes on a schedule, loops through tool calls until it stops, and waits for a human at the decisions that carry liability. Howie Liu, founder of Airtable HyperAgent, presented that pattern at the AI Engineer World's Fair 2026 alongside a live startup competition with $100,000 in prizes.

Hiring employable agents: the five form factors Liu described

Hiring employable agents is the final step in a progression Howie Liu has tracked through five product form factors: completions, chatbots, agents, always-on agents, and agents that orchestrate other agents. Liu, who founded Airtable, described that ladder at the AI Engineer World's Fair 2026 in San Francisco.

The ladder matters because each rung changes what the human does. Completions ask a model to continue a token string until it hits a stop token. Instruction-tuned models enabled the chatbot. An agent, in Liu's definition, is a model that recurses on itself with an open-ended tool set rather than following a fixed workflow.

The fourth rung is the one Liu calls an always-on agent. It wakes on a heartbeat mechanism, restarts after reaching a stopping point, and keeps moving toward a goal without a fresh human prompt. The fifth rung hands work to other agents, so one agent manages several scoped tasks in parallel.

Liu attributes the shift to model coherence on longer, more open-ended problems. His comparison points are BabyAGI and AutoGPT, which he calls science experiments that eventually drifted into rabbit holes. Current models, in his account, hold a long job together well enough to hand pieces between agents.

One caveat sits on the ladder. Liu presents it as a product-form-factor reading from a vendor stage, not an independent benchmark. HyperAgent is his own product inside Airtable the model names he offered as examples on stage should be treated as his listing, not as verified specifications.

What separates an always-on agent with a heartbeat from a chat agent

An always-on agent keeps working after the task that a chat agent would have closed. The heartbeat mechanism wakes the agent on its own schedule, and each wake-up runs many turns: a tool call returns, the model runs again, and the loop repeats until it reaches a stopping point that can itself be restarted.

Liu credits OpenClaw with popularizing the concept and says HyperAgent now ships the same mechanism. The practical difference is what happens when the agent gets stuck. A reactive agent waits for the user to notice. A heartbeat agent gets another turn to find a way around the blocker on its own.

That has a cost side that Liu does not quantify on stage. Every wake-up consumes inference, so an always-on agent spends tokens while nobody is watching. HyperAgent sells inference credit rather than seats, which fits a product where the meter runs during idle hours.

The heartbeat also changes review work. A person supervising five reactive agents answers five prompts. A person supervising five heartbeat agents reads a status view of what each one is chasing and what each one is blocked on, and intervenes there.

How the HyperAgent landscaping example runs end to end

A landscaping company runs its client pipeline on HyperAgent in the example Liu walked through on stage. A homeowner submits an inquiry with photos of a messy backyard and asks for a quote and a proposal. That inquiry routes into a triage agent.

The triage agent checks the scope and the quality of the lead, does some research, and hands the work to a surveyor agent. The surveyor inspects the imagery the submitter provided and drafts what a relandscaped version could look like.

Liu said each agent runs in a full sandbox VM, so it can write code, manipulate files, and call tools such as ffmpeg to pull individual screenshots out of a video. The proposal that comes back is an interactive web page with mocked-up imagery and a quote based on the surveyed factors.

The human owner then approves sending the pitch. Liu framed that approval step as the manager's job, not a failure of automation. A binding quote and a commitment to do the work are the two points where a person still has to sign off in his account.

The business detail Liu highlighted is the deal size, not the technology. A high-touch proposal with custom imagery used to be economical only for clients paying millions, in his telling. Here it is generated for a small backyard job worth around $10,000.

Situational awareness and feedback that updates the agent

Situational awareness is what lets one agent know which job it is on and what has already happened. Liu describes agents that accumulate memories and skill updates from every interaction instead of waiting for a fine-tuning run on collected data.

In the landscaping example, the owner replies to the agent through Slack or email the way they would to a new hire: remember this cost lever, format the proposal this way next time. The agent updates from that correction rather than from a retraining cycle.

Liu ties the point to a broader argument about context that persists across interactions. His claim is that an agent with a static prompt and fixed skills stalls, while an agent that folds in real-world feedback keeps improving at the job it already holds.

This is a design claim from the person selling the product, and the transcript contains no measured retention or accuracy number behind it. It is worth separating from the mechanism, which is straightforward: store the correction, retrieve it on the next run, apply it before the user has to repeat themselves.

What the human still owns once agents run the work

The human owns decisions with consequences, not the labor that produces them. Liu's landscaping example puts two of those decisions in human hands: approving the quote that goes out and agreeing to take on the work.

Everything before those points can run without a person. Everything after depends on a person having said yes. Liu uses the phrase unblocking agents to describe the job, and compares the role to a manager reading upward reports from a team.

He also describes a shift in the interface toward an orchestration control plane, a single view of what every agent is working on, what each one is blocked on, and which agent handed work to which other agent. The person zooms out to the whole board instead of into one task.

There is a friction the transcript does not resolve. An approval step is only as good as the reviewer's ability to audit a proposal an agent produced. Liu does not describe how a small-business owner verifies a generated quote before signing it, and that gap is where mistakes would surface.

Inside the Startup Battlefield at the World's Fair

Three agent-native startups pitched live for $100,000 in prizes, drawn from a founding cohort HyperAgent calls the 500. The judges were Howie Liu, Theo Browne of T3 Chat, and Joshua Xu of HeyGen, and the event was recorded at the AI Engineer World's Fair 2026.

Comment IO took the grand prize and the $50,000 first-place award. Built by Foundry placed second for $30,000, and Kamad placed third for $20,000. The organizer framed the contest as a test of why an idea matters, not as an investment pitch.

The three finalists sat in different parts of the market, which is what made the judging interesting. One sells deep domain orchestration, one sells a shared document surface, and one sells a product team for creators.

Startup	Pitch	Judges' pushback	Outcome
Kamad	Agentic execution layer for physical commodity trade, built on a patent-filed state machine the founders call the CFC	How to stay at the frontier as models improve and rivals ship weekly	Third place, $20,000
Comment IO	Multiplayer markdown editor built for people and agents to share context in the same document	Why agents cannot simply collaborate inside Google Docs	Grand prize, $50,000
Built by Foundry	Teams of agents that build recurring-revenue products for content creators	Whether creator businesses actually work outside a few niches	Runner-up, $30,000

Kamad, Comment IO and Built by Foundry in the judges' crossfire

The three finalists answered different objections, and the objections reveal what the judges thought was hard about each shape of business.

Kamad: deep domain versus a smarter model

Kamad's founders described an agentic layer for physical commodity trade, where agents screen both sides of a deal for identity and sanctions, catch document fraud, and advance through a state machine only when every deny condition clears. The judges asked whether a sufficiently capable model at Anthropic OpenAI could absorb that work. The founders answered with proprietary data, domain expertise, and deal sizes where one transaction can run to $200 million or more.

Comment IO: why not just use Google Docs

Comment IO is a multiplayer markdown editor for people and agents to work in the same document. Joshua Xu asked whether the format should be HTML rather than markdown. Liu asked why agents cannot just collaborate in a shared Google doc. The team's answer was that existing writing tools expose an MCP for reading and writing but do not let people and agents edit together in real time, and that enterprises want an open, pluggable format.

Built by Foundry: creators as founders

Built by Foundry deploys agent teams that build recurring-revenue products for content creators. Theo Browne pushed back hard, saying he had tried the creator-as-founder path and that creators get more from a wider set of sponsorships than from launching their own company. Weston Belgettes answered that tech is a hard niche and cited a homesteading creator whose audience pays for a tool that tracks freeze-dryer settings.

The disagreement is the most useful part of the session. Browne's objection is an operator's experience claim, not a measurement, and Foundry's answer is a customer example rather than a retention figure. Neither side settled it on stage.

FAQ

• What are employable agents? Employable agents are AI workers you delegate a job to rather than a prompt. In Howie Liu's framing at the AI Engineer World's Fair 2026, they run on a heartbeat, accumulate memory from feedback, hand work to other agents, and escalate binding decisions to a human.

• What is HyperAgent? HyperAgent is an agent product inside Airtable, the spreadsheet-database platform Liu co-founded. HyperAgent runs agents in sandbox virtual machines, supports always-on execution through a heartbeat mechanism, and is sold with inference credit rather than per-seat pricing.

• How is an agent different from a workflow? A workflow follows a scripted sequence of steps. An agent, using the definition Liu borrows from Anthropic a model that recurses on itself with an open-ended set of tools and reasoning steps, so it can choose its own path instead of following a fixed order.

• Does HyperAgent let agents spark other agents? Liu describes a tier where one agent orchestrates several others, fanning out work the way he says he uses workflows in Claude Code. He associates that pattern with the newest model generation rather than with earlier agent frameworks.

• Who won the HyperAgent Startup Battlefield? Comment IO won the grand prize of $50,000, Built by Foundry was runner-up with $30,000, and Kamad placed third with $20,000. The three finalists came from a founding cohort HyperAgent calls the 500.

• Can agents run a whole business without a person? Not in the example Liu presented. The landscaping business keeps a human at the approval step, where a binding quote and a commitment to do the work get signed off, and describes everything else as agent-executed.

• What did Theo Browne object to in the creator pitch? Browne said creator-owned companies burn creators out and that creators earn more from a wide sponsorship set than from launching their own product. He framed his objection as experience, not as data, and the exchange ended without a resolution.

• Is Skala Blog involved in this event? No. Skala Blog is a separate tool that turns YouTube videos into written articles, and it is mentioned here only as the production route for a transcript like this one.

• Where did this event take place? The AI Engineer World's Fair 2026 ran in San Francisco, and the session combined Liu's keynote with three startup pitches recorded in the same block. The founding-500 cohort and the $100,000 prize pool were the organizer's framing for the competition.

Source video