I think we’re onto something

We started this year at $0 in revenue. Today, we’re at multiple millions in committed revenue across several seven-figure contracts. Not run-rate, not pipeline, not “contracted” in the way people stretch that word. Hard commits.

They say if something happens once, you got lucky. If it happens twice, it’s a coincidence. If you’ve done it three times, you might be onto something. We think we’re onto something.

And we’ve raised $15M from Matrix and Lachy Groom to accelerate it.

Over the last 6 months, Takeoff agents have conducted 3.5 million interactions for customers across 5 industries spanning lending to healthcare. Takeoff agents have beaten competing agents in head to head tests and consistently outperformed the existing processes, systems, and people in place before them.

Takeoff is a long horizon agent runtime, a harness for agents that operate like humans but scale like software. They aren’t AI teammates that live in Slack, it’s not an agent framework, and we don’t sell seats.

Our customers send an API request for a job to be done (originate this loan, schedule this patient, call insurance for a prior authorization, activate this user), and Takeoff agents complete that job across APIs, voice, SMS, email, and whatever else it takes. The job might span hours to weeks of work, and we deliver progress updates and the final outcome back to our customer via webhooks.

Our customers never build, configure, or think about the agents. The agents do the work and get better at it week over week, just like how your best employees do.

Our agents are priced on outcomes. A customer doesn’t commit millions in spend unless that spend is making them tens of millions, and the way an agent makes them tens of millions is by doing the work their revenue comes from. So the unit we sell is a unit of revenue-aligned work. When the customer grows there’s more work, and when there’s more work the bill grows with it. We never have to sell them a bigger contract, because their growth does it for us. One day they look up and they’re happily spending millions with us.

There are two kinds of AI agents, and they need different harnesses.

The first is the human-in-the-loop agent, the kind that makes a knowledge worker more productive. Claude Code, Codex, Cowork, and the vertical tools have swept through the industry and let people do 10x the work they could before. Some of them run for hours to days. But every run is started, steered, and accepted by the person at the keyboard.

The second is a fully autonomous agent. An agent that doesn’t need you to tell it what to do. Nobody steers a Waymo. You sit back and it drives. The industry has witnessed it with customer support, where an agent handles a growing share of requests end to end with no human in the middle. As harness engineering advances and the models get better, the set of jobs that can be done this way keeps widening.

The human-in-the-loop harness has a standard shape by now: a frontier model, a sandbox, tools, skills, and connectors to your data. That harness is the right answer for a lot of work. It is the wrong answer for anything that has to happen at the speed and scale of a customer-facing operation, where the agent has to be reachable by the outside world and accountable to an outcome.

Our harness is for jobs that run end to end. A great model, tools, skills, and connectors are necessary but insufficient. On top of them we run orchestration across voice, SMS, email, and fax, persistent memory for each lead across interactions that span days and weeks, compliance guardrails, and enough determinism that you can predict how the agents will behave.

Long horizon doesn’t have to mean long running. A loan origination is long horizon because it spans 45 days and dozens of decision points, not because an agent burns tokens for 45 days straight. Most of the highest-value work people do in regulated industries looks like that. It takes persistence, judgment, and the ability to be summoned by the world, not a sandbox and a filesystem.

The bitter-lesson pilled reader will point out that a frontier model and its harness can, in principle, do any kind of work. That’s true. Economists have a standard story about exactly this:

What took forty years wasn’t the electricity, it was rearranging the factory around it. A frontier model in a general-purpose harness is one big motor in a room full of belts. The models are electricity, and at Takeoff, we build the motors.

We built Takeoff on the premise that an enterprise AI agent should be aligned with the revenue it’s supposed to generate. Token-based AI businesses have incentives that rhyme with attention advertising. The more time you spend, the better; the more tokens you burn, the better. This business model pays the platform for the work it makes, not the work it finishes.

We aren’t trying to maximize how many hours our agents run so we can pass you the bill. We care about whether the job gets done. If a mortgage loan isn’t funded, we don’t get paid. If a healthcare team’s prior authorizations don’t get approved, we don’t get paid.

Our revenue is a direct function of our customers’ results, so every improvement we make directly translates into more revenue for them, and then for us. And it compounds: every agent has a measurable goal it’s graded on, and the platform learns from every interaction for continuous improvement.

Daniel, the company is Takeoff.