When we started Airtable 12 years ago, nearly every investor we talked to told us the same thing: you can't build everything for everyone. Too many use cases, too much surface area. We proved them wrong. We built a horizontal platform that runs business processes across retail, consumer, financial services, media, and more.
We're now applying that same philosophy to agents.
The next form factor for agents
Consider the dominant product form factors of the last few years. We started with completions and chat, then levelled up to agents. At first, agents needed to be prompted. Now agents wake up on their own to complete work proactively.
Progressing to a Fleet of agents.
What comes next? We believe the next product form factor is the orchestration of a fleet: coordinating complex work across many agents acting as part of an organization. Agent-to-agent coordination, aimed at longer and harder jobs.
Models are now good enough over long, open-ended horizons that you can start applying real organizational principles: break a big job into differently scoped tasks, give each one to a different agent, and let them hand work back and forth.
Real example: Landscaping company
Some of the most creative fleets we see on Hyperagent are coming from traditionally offline businesses. Here's one based on a real customer in landscaping.
First, a client submits an inquiry with photos of a messy backyard, some basic info, a request for a quote and a proposal.
Inbound lead processed by agents.
The fleet wakes itself
The inquiry lands with Sage, the orchestrator. Sage reads the intake and does the first pass a good office manager would: what's the scope, is this a real project, is this a quality lead worth pursuing?
Sage decides it's worth taking on. It hands the job to Surveyor, a specialized agent built for design and quoting, and briefs it on what the client asked for.
A simple agent org structure.
You're not building one agent that does everything. You're building an org where a coordinator routes work to the specialist best suited to it, the same way you'd assign a job to the right person on your team.
Surveyor gets to work. Every thread on Hyperagent runs in a fully capable sandbox VM, so the agent can write code, manipulate files, and reach for real tools. It uses something like ffmpeg to pull individual frames out of the client's video, studies the space, and assembles a high-touch proposal: mocked-up imagery of the redesigned yard, a real pitch, a real quote.
Quoting & design agent running its own sandbox
A proposal like that used to be something only a high-end designer or landscaper could produce, and only for a multimillion-dollar client. Now a small business can send that same quality of pitch for a $10,000 backyard job.
The fleet prompts humans
Once Surveyor has a pitch and quote ready, the work comes back to a human for review. Sarah Guo wrote about this recently, and we think she's right: the real bottleneck in deploying useful agents isn't model capability anymore, it's the human layer that owns policy and gates high-stakes decisions. Sending a binding quote to a customer is a perfect example.
Agent shares the proposed landscaping transformation with business owner for review.
So Surveyor's work routes back through Sage, and Sage messages the business owner directly, the way a good employee reports up. “Here's the lead, here's the survey, here's a full pitch deck and proposal, here's a quote I built from everything I found. Approve it and I'll send it.”
The proposal renders as an interactive web page with a before/after transformation, the kind of thing that would have been completely out of reach a year ago for a landscaper quoting a $10,000 job.
The fleet learns about your org
Useful agents have to learn in real time, not on the schedule of your next fine-tuning run. Feedback get crystallized into memories and skills, compounding continuously.
In practice that looks like ordinary back-and-forth with a teammate, in Slack or email. The fleet accumulates more internal context and expertise with each run.
Agents learning from feedback delivered via Slack.
The fleet coordinates without you
Once you have several capable agents that can also talk to each other, a lot of the coordination stops running through you. Put them in a shared channel, a group chat in Slack, and they hand work back and forth directly. Sage flags something for Dispatcher, they settle it between themselves and keep moving. You can watch the whole thing happen, and you can jump in the moment you want to, but most of the time you don't need to. The work closes without waiting on you to relay a message from one agent to the next.
You go from being the wire that connects every agent to being the person who reads the thread when something actually needs a decision.
Overseeing your fleet
Once agents are individually capable and coordinating with each other, the human's job shifts almost entirely to unblocking them. Coding moved through this arc, and we're about to see it happen for all knowledge work.
First, solo coding: single-threaded, one file, one problem, one person. Then GitHub Copilot's early autocomplete, which was really the completions form factor applied to code. Then Cursor's original chat experience, where you could talk to an agent and it made more complex edits. Now the best agentic developers I know spend most of their time overseeing a fleet, and going to sleep without setting the agents loose on something overnight feels like leaving your whole team idle.
Agents move their own work forward on a kanban.
That changes the interface you need. You need a control plane: one place to see, at a glance, what every agent is working on, what it's blocked on, and who's handing off to whom. The job becomes zooming out to see the whole system at once, a kind of SimCity view across everything running, instead of zooming into any single task.
Your new job
Agents are joining your org chart. The most important skill to learn right now is how to oversee and lead a fleet of agents. Design the agent org structure, cultivate the context layer, make every interaction legible to them, and define when you actually need to be in the loop.
