AGI is here. Now harness it.
Introducing Hyperagent, a new product from the team at Airtable.
I’ve been burning through billions of tokens a week. Not theorizing about where agents are going, but building with them, and building them, in the most hands-on way possible.
The models didn’t just get better. They crossed a threshold — from turn-by-turn assistance to agents that run autonomously on open-ended problems for hours. In software engineering, where this shift is most visible, the best engineers went from mostly manual coding to mostly agent coding in a matter of weeks. The team building Claude Code writes 100% of their code through AI. Traditional two-week sprint cycles are compressing into hours. I’ve done it myself. OpenClaw has shown what happens when you give models a browser, a shell, and real tools — and captured massive mainstream attention doing it. We are, for practical purposes, in the era of artificial general intelligence. And this shift can’t be understood from a distance.
Software engineering is where this is playing out first. Every other domain of knowledge work is next.
Intelligence is the prerequisite, not the product
Intelligence is no longer the bottleneck. Harnessing it is. And the gap between the teams who’ve figured that out and everyone else is already enormous.
Moore’s Law compounded computing power for decades. AI compute is compounding at a far faster rate, and every generation of models is converting that power into dramatically more capable intelligence.
But the breakthrough processors of the personal computing era would have remained novelty marvels of circuitry without the work Apple and Microsoft did to build the entire computing experience around them. And each successive generation of hardware didn’t just get faster — it enabled and demanded entirely new product forms. Mainframes gave way to PCs, PCs to mobile, each unlocking capabilities the prior form factor couldn’t support.
The same must be true for models. We’ve moved from simple chatbots to copilots to what’s possible now: autonomous agents that reason, act, and compound their work over time. The intelligence is there — but it’s Einstein in a Faraday cage until you build the right system around it.
Closing that gap — between what powerful technology can do and who can actually use it — is in Airtable’s DNA. It’s what we did for app creation. And it’s what this moment demands again. OpenClaw, impressive as a demonstration of unleashed state-of-the-art agent capability, is still a bleeding-edge power-user tool — requiring dedicated hardware, carrying real security risks, and far from ready for everyday use inside a business. The gap between the bleeding edge and everyone else is growing fast. The models will keep getting smarter on their own. What matters now is the system that lets agents learn, compound, and scale.
What we’re building
What’s missing isn’t any single feature — it’s the complete system. The orchestration, the tools, the memory, the skills, the ability to deploy and manage agents across an entire organization. No one has built it yet. We are. We’re calling it Hyperagent.
When the cost of intelligence drops to zero — and it is dropping, fast — the organizations that have this system will compound every interaction, every correction, every workflow into something that makes the entire operation more capable over time. The ones that don’t won’t keep up.
State of the art agents with complete, safe, computing environments and powerful tools at their disposal
At the core of any agent system is the agent itself. Before an agent can learn or compound, it needs to be genuinely capable of doing real work — not answering questions about work, but doing it.
Our agents have the full toolkit at their disposal: a browser, a full computing environment (an isolated machine per session — not a sandbox, a real computer with a filesystem and shell), image and video and audio generation, hundreds of enterprise integrations connected through in-chat OAuth, data warehouse access, geolocation and mapping tools, and deep research capabilities that synthesize across dozens of sources. But having tools isn’t the point. Knowing how to use them together — that’s what separates a demo from real work.
Say you’re an interior designer doing personalized outreach to homeowners. You give the agent a neighborhood. It browses Zillow to pull recent listings with interior photos, navigates to Google Street View to capture exterior shots of each property, and analyzes each home’s existing aesthetic. Then — and this is where it gets interesting — it takes the actual interior photos and uses them as seed images for Nano Banana Pro to generate redesigned rooms that match each home’s specific style. It feeds those redesigned stills into Veo 3 as first-frame images to produce walkthrough videos of the reimagined spaces. And it compiles everything into a publishable webpage with an embedded interactive map of nearby amenities — coffee shops, restaurants, parks. The agent didn’t just use six tools sequentially. It chained them creatively — each output becoming the input to the next — and produced a finished deliverable. That’s what separates a demo from real work.
Unlock domain expertise with skill learning
Our agents retain your preferences, constraints, and context across sessions. But memory alone isn’t enough.
Skills are what turn a general-purpose agent into a domain expert. Tell the agent to learn the Stripe API, and it autonomously finds documentation, identifies key endpoints, generates working scripts, tests them, and packages everything into a reusable skill. But skills aren’t limited to APIs — your due diligence framework, your press release methodology, how your team formats cohort retention analysis. These are process skills that make the difference between generic output and output that reflects real expertise.
Skills improve with every run. You spend a session teaching the agent how your firm evaluates early-stage companies — how you identify the moment of inflection where a market tips from nascent to inevitable, the founder qualities that signal someone who can ride that wave, how you assess whether a technical architecture becomes a compounding moat or gets commoditized within two years, and the format your partners need to make a conviction call in the Monday meeting. Now any partner at the firm says “pull together a view on this company” and gets analysis that reflects your firm’s thesis — not a generic VC template anyone could pull off the internet.
After every conversation, the system generates improvement suggestions — new skills, memories worth retaining, refinements to the agent’s instructions. These can be reviewed manually or auto-accepted, so the agent silently accumulates knowledge with every interaction without requiring explicit training.
One click to deploy as intelligent coworkers for your entire org
An agent with skills and memory shouldn’t only be useful to the person who built it.
We’ve built the ability to package everything — skills, memories, a system prompt, tool access, model selection, budget limits — into a role-based agent with a clear purpose. The “Customer Intelligence” agent. The “Content Production” agent. The “Competitive Research” agent. Each one carries the accumulated knowledge and capabilities relevant to its domain, available to anyone in the organization.
Deploy any agent to Slack with one click. But these aren’t the chatbots you’re used to — the ones that sit inert until someone @mentions them with a perfectly formatted request. Hyperagent agents are genuine participants. They read the room. They follow conversations across channels, understand context and nuance, and engage when they have something valuable to contribute — the way a great colleague would. A rep mentions an upcoming account meeting in passing. The agent recognizes the signal, pulls the latest activity from Salesforce, reviews recent call transcripts, cross-references the internal knowledge base, and drops a competitive positioning brief into the thread before the meeting starts. No one asked it to. No one @mentioned it. It just understood the context and acted.
This is how agents go from something one person experiments with to infrastructure an entire organization compounds on.
Orchestrate, monitor, and continuously improve an entire fleet of agents
Once agents are deployed across an organization — handling real work, invoked by people who didn’t build them — managing them is where the real leverage lives.
You wouldn’t hire a team and never check their work. The right architecture treats every agent run as a data point. Every run gets scored for quality automatically, using configurable evaluation rubrics with an independent judge model — not the agent grading its own work. Start with general rubrics. Get more precise as you invest in defining what “good” looks like for each specific agent.
From there, you have visibility across your entire fleet. Where agents are scoring well, where they’re falling short, what patterns emerge across thousands of runs. A/B test any change against real workloads before rolling it out. Swap models to optimize cost or quality. Adjust instructions and measure the impact with hard numbers, not vibes.
The loop closes: agents do work, work gets scored, patterns surface, the system suggests improvements, improvements get validated against real data, validated changes get deployed, agents do better work. Continuously, at scale.
This is the operating system for the age of AGI.
Hyperagent is already real and working. We’re testing it with a small group of early users, and we’re looking forward to onboarding more people in the coming weeks.
A few weeks ago we launched Superagent — our first standalone agent product, focused on research use cases. Hyperagent is the complete vision: a full platform for building, deploying, and managing agents that don’t just research, but act across every tool and workflow your organization runs on.
Airtable is growing at scale, generating cash and extremely well capitalized, with a distribution base of hundreds of thousands of organizations orchestrating strategic operations at the top companies in the world. We power end-to-end product operations in the largest tech organizations and leading frontier labs, content production operations at the majority of top media companies, and end-to-end marketing campaign orchestration at the largest brand companies. The builder skillset and community for building and deploying apps into an organization is highly relevant to building and deploying frontier agents.
The products are natural complements. The hard part of human-agent collaboration isn’t intelligence — it’s coordination. That requires a shared system of record with full visibility into the work being done — and that’s exactly what Airtable already is for hundreds of thousands of teams. Hyperagent delivers the frontier agents platform that complements that data layer, but also serves standalone use cases.
Go to hyperagent.com to join the waitlist. We can’t wait to see what you build.
Heck yes Howie. Love it. Zig when everyone else is zagging. Writing history. Ebb and flow. When people are building buildings, build companies that build new tools in new ways to build new cities! Can’t wait to try it out. Have been loving seeing Superagent come to life and in the real world. Keep burning tokens and burn the ships!
Would love to test it! Already signed up!