Hyperagent: Cloud AI Agents Every Business Owner Should Test

Six months ago I wrote a deployment guide for Open Claw and a security guide to go with it. Both are still among the most-read things on this site. Both are also, quietly, a confession: the most capable personal AI agent available to a business owner in early 2026 required a VPS, a reverse proxy, a hand-edited config file, and a genuine understanding of prompt injection before it was safe to point at your inbox.

That was the deal. Enormous capability, and a setup process that filtered out roughly everyone who actually runs a company.

Hyperagent is the first serious attempt to remove that filter. It is a cloud agent platform where every agent gets its own isolated computer, persistent memory, a set of learned skills, and a spending cap, and you configure the whole thing by talking to it. Howie Liu, who built it, put the positioning about as bluntly as it can be put: "Hyperagent is to Open Claw what a Macintosh is to a Linux box."

I think that comparison is close to right, and I think it matters more for operators than for engineers. Here is what Hyperagent actually is, how the architecture works, what it costs, where it breaks, and what a business owner should do about it in the next 30 days.

Where It Came From

Hyperagent was built by Howie Liu and the Airtable team. It started life as Superagent, which launched on January 27, 2026 as Airtable's first standalone product in thirteen years. Hyperagent followed in February as the bigger idea, Superagent was discontinued and rebuilt into it, and public signups opened in April 2026.

One more piece of context, because it explains why this product suddenly has everyone's attention: in August 2026, Bending Spoons agreed to acquire Airtable for $2.25 billion, which nets out to roughly $1.285 billion in enterprise value against Airtable's cash. Hyperagent was carved out into its own company before the deal closed. Liu went with it. The agent platform was the part worth keeping.

What Hyperagent Actually Is

Strip away the marketing and Hyperagent is three things stacked together.

1. A cloud computer per agent session. Not a chat window with plugins bolted on. Every session spins up an isolated environment with a real browser, a shell, a file system, code execution, image and video generation, and connections to hundreds of third-party tools. The agent can log into a site, scrape a table, run a Python script against it, generate a chart, and publish the result. Nothing about that requires a machine on your desk.

2. A fleet, not an assistant. You do not build one agent. You build a roster. One does research, one does outreach, one does content, one reconciles invoices. Each carries its own tool access, its own memory, and its own budget cap. That last item is the part most people skip past and shouldn't.

3. A learning layer. Skills are reusable methods you teach once and any agent can pick up. Memories persist across every conversation, scoped globally or per agent. After each run, the system proposes new skills, new memories, and adjustments to tool access. The agent you have in month three is measurably not the agent you hired in week one.

Sitting on top is a command center: cost per run, output quality scored against custom rubrics, and A/B testing across agent versions. That is the piece that makes it a platform instead of a toy.

How Agents Get Triggered

An agent that only runs when you open a tab is a chatbot with extra steps. Hyperagent agents fire from a thread, a schedule, an inbound email, a Slack mention, Telegram, a webhook, MCP, or a raw API call.

The Slack deployment is the one I would push most operators toward first. You drop the agent into a channel and your team @-mentions it the way they would mention a coworker. No new tool for anyone to learn, no adoption curve, no license nobody logs into. If you want the deeper background on how agents talk to your existing tools, our MCP servers guide covers the protocol underneath most of this.

Why the Cloud Architecture Is the Whole Story

Here is where the Open Claw comparison earns its keep, and why I think this is a genuine category shift rather than a new logo.

Open Claw is extraordinary software. It is also software that, by design, runs on your machine with access to your files, your shell, and your credentials. When it went viral, security researchers found more than 1,800 Open Claw dashboards reachable on the open internet without any authentication inside 48 hours. The default config bound the gateway to all interfaces. The skills marketplace had no code signing and no review process. Every one of those is fixable, and our security guide walks through fixing them, but "fixable" and "fixed" are different words, and the gap between them is where businesses get hurt.

Hyperagent's answer is architectural rather than procedural. As Santiago Valdarrama put it: you do not need a Mac Mini or a server, every agent runs in the cloud in its own isolated environment. The isolation is not a setting you remember to turn on. It is the only mode.

For a business owner, that changes the risk calculus in four concrete ways:

• Blast radius is bounded by default. A compromised session owns a disposable container, not your laptop and not your network.
• Credentials live in a managed integration layer, not a plaintext config file you edited at 11pm.
• Spend is capped per agent. An agent stuck in a loop burns its own budget and stops, instead of quietly running up a model bill.
• There is an audit trail. Every run is a thread with the full history of what the agent did, which is the difference between "the AI did something" and something you can actually review.

If you already run Open Claw and like it, Hyperagent's onboarding includes an import path directly from it. That is a deliberate and fairly confident move on their part.

What This Replaces (and What It Doesn't)

Operators keep asking me whether this kills their automation stack. Mostly no, and the distinction matters for budgeting.

Traditional automation (Zapier, Make, n8n) executes steps you defined. It is deterministic, cheap per run, and brittle: change the input shape and the scenario breaks. It is still the correct tool for high-volume, well-defined, must-never-vary work. Our Make vs n8n vs Claude Code comparison breaks down where each still wins.

Chatbots answer questions. They do not act.

Hyperagent-style cloud agents take an objective, plan the approach, use real tools, recover from failure, and deliver an artifact. Expensive per run, resilient to messy inputs, and genuinely capable of judgment calls.

The right mental model is not replacement, it is tiering. Deterministic automation for the high-volume plumbing. Cloud agents for the work that used to require a person to look at it and decide. If that framing is new to you, our autonomous AI agents guide covers the taxonomy in more depth.

Where Business Owners Should Actually Point It

I want to be specific here, because "deploy AI agents" is advice that helps nobody. These are the deployments I would rank first for a company doing $1M to $20M, in order of how quickly they pay for themselves.

Inbound lead research and enrichment. A form fills, the agent opens the company site, pulls headcount and tech stack, checks recent news, scores the lead against your ICP, and posts a briefing to Slack before your rep opens the tab. This is the single best first agent because the output is obviously right or obviously wrong within a day.

Weekly competitive and market monitoring. Scheduled agent, runs Monday at 6am, checks competitor pricing pages, job postings, ad libraries, and review sites, and delivers a diff against last week. Nobody on your team was doing this consistently. I promise.

Proposal and quote assembly. The agent pulls the discovery notes, applies your pricing rules, drafts the scope, and hands a human the last mile. Keep the human on the last mile.

Recurring reporting. Pull from your warehouse or your CRM, compute the numbers, write the narrative, publish. This is the workflow I have seen collapse the most hours, because "assemble the monthly report" is usually two days of a good operator's month.

Support triage and drafting. Classify, retrieve the relevant history, draft the reply, route the ones that need judgment. Draft, do not send, for at least the first month.

Vendor and invoice reconciliation. Cross-check invoices against contracts and flag the variances. Boring, high-yield, and the kind of thing that never gets staffed.

Notice what is not on that list: anything customer-facing and unreviewed, anything that moves money, and anything where a confident wrong answer costs more than a slow right one. Those come later, behind approval gates, or never.

What It Costs, Honestly

Hyperagent does not publish a pricing page. That is worth sitting with for a second if you are the person who signs the invoices.

So here is my own account instead, which is worth more than anything I can quote secondhand. Over the last 30 days I ran 422 million tokens across every thread and agent and spent $385.67. That lands right around $0.91 per million tokens. The unit rate is cheap.

Volume is what gets you. Those 422 million tokens came from only 8 active days, and the heaviest single day burned 103 million on its own, which at that rate is roughly $94 in an afternoon. Hyperagent is token-hungry in a way that catches out anyone used to reasoning about cost per API request. You are not buying requests, you are buying an agent that reads, retries, and re-reads until it is done.

The pattern worth planning around is that the burn is front-loaded. Standing an agent up costs considerably more than running it, because the build phase is where you iterate on instructions, rerun the thing, and correct it. Steady-state operation settles well below that. Budget for a spike while you are building and do not extrapolate that week across the year.

Treat those numbers as one account's experience rather than a rate card, and note that the spend and token figures come from slightly different views in the dashboard. The order of magnitude is the useful part.

On the published side: billing is consumption-based and bought as credits, subscriptions have been reported starting around $20 per month, and there is a pay-as-you-go option with no monthly commitment. New accounts have been getting a $1,000 credit bonus. Liu has also put $10 million in inference credits behind a "Founding 500" program, offering $20,000 in credits to each of 500 founders building agent-first companies.

The credit-burn complaints are real and you should plan for them. The most consistent one: web research is the single most expensive action on the platform, which is awkward given how many agent workflows start with a search. The practical workaround operators have landed on is to do cheap retrieval elsewhere and hand the agent a pre-qualified shortlist to work from. Users have also reported being unable to stop a run mid-flight while watching the spend climb, which is exactly when you most want to stop it. Use Plan mode for anything whose cost you cannot predict: review the plan first, then execute.

This is what per-agent budget caps are for. Set them low, watch the command center for two weeks, and raise them deliberately.

The 30-Day Rollout I Would Run

If you are going to test this, test it like an operator instead of a hobbyist.

Week 1: One agent, one workflow, one owner. Pick the highest-volume research or reporting task in your business. Assign one person to own the agent. Set a hard budget cap. Ship nothing to customers.

Week 2: Write the rubric before you scale. Define what "good output" means in writing, then use the built-in scoring to grade the runs against it. This is the step everyone skips, and it is the only reason you will be able to tell in week 6 whether this is working or whether it just feels impressive.

Week 3: Codify what worked into skills. The runs that produced good output contain a repeatable method. Save it as a skill. This is the compounding mechanism and it is the actual product.

Week 4: Add the second agent and connect Slack. Only now. A second agent before the first one has a rubric is how you end up with an expensive mess and no way to diagnose it.

At the end of 30 days you should be able to answer one question with a number: what did this cost per completed unit of work, versus what it cost before. If you cannot answer that, the pilot was not run correctly.

The Honest Caveats

I am recommending you test this. I am not recommending you bet the quarter on it.

Hyperagent is early. Public production references are thin, pricing is not published, and the company just spun out of a business that was itself sold at roughly 80% off its 2021 peak valuation. None of that makes the technology worse, but it does mean you should treat it as a capable vendor you are piloting, not infrastructure you are standardizing on. Keep your data exportable. Keep your critical workflows portable. Do not build something you cannot rebuild.

The governance question is also unsettled across this entire category. An agent with browser access, shell access, and your credentials is a powerful employee with no HR file. Approval gates, scoped tool access, and per-agent budget caps are not paranoia, they are the minimum viable controls. Use all three.

And the durable point underneath all of this: the platform is not the advantage. Every competitor can sign up for the same product this afternoon. The advantage is the skills and memory you accumulate, which encode how your business actually operates. That asset takes months to build and it is the only part a competitor cannot buy.

Where This Fits With Custom Builds

For a lot of companies, a platform like Hyperagent is the right answer and the story ends there. Sign up, run the 30 days, keep what works.

It stops being the right answer at a predictable point: when the workflow needs to live inside your product, when compliance requires the data never leave your environment, or when per-run economics stop working at your volume. That is when you move the proven workflow onto your own stack, which is a much easier project once an agent has already proven the workflow is worth having. Piloting on a platform first is the cheapest possible way to find out what to build.

That transition is most of what we do at White Space. If you want a second opinion on which of your workflows belong on a platform and which belong in your own infrastructure, get in touch.

Frequently Asked Questions

Who makes Hyperagent? Howie Liu and the Airtable team built it. It launched as Superagent in January 2026, was rebuilt into Hyperagent that spring, and was spun out as an independent company in August 2026 ahead of Bending Spoons acquiring Airtable.

How is Hyperagent different from Open Claw? Open Claw runs on hardware you own and manage, which gives you total control and total responsibility for security, updates, and isolation. Hyperagent runs every agent in an isolated cloud environment with managed credentials and per-agent budget caps. You trade some control for a dramatically shorter path to safe production use. Hyperagent also offers a direct import from Open Claw.

What are subagents and why do they matter? A subagent is a specialized agent that a parent agent delegates a focused task to. It runs independently and returns only its result. This keeps the parent agent's context clean and lets you assign narrow tool access and separate budgets to each specialist, which is both cheaper and safer than one agent with access to everything.

What does Hyperagent cost? There is no public pricing page. Billing is consumption-based and sold as credits, with subscriptions reported to start around $20 per month and a pay-as-you-go option alongside them. On my own account the effective rate worked out to about $0.91 per million tokens, but I burned 422 million tokens in 30 days, so the cheap unit rate matters less than the volume. Expect the heaviest spend while you are building an agent rather than running it.

Should a small business use this instead of Zapier or Make? Not instead of. Deterministic automation is still cheaper and more reliable for high-volume, well-defined tasks. Cloud agents are for the work that previously required a person to look at it and make a judgment call. Most businesses should run both.

What is the safest first workflow to hand an agent? Internal research or recurring reporting. Both produce output a human reviews before anyone outside the company sees it, which means a bad run costs you an hour instead of a customer.