0:00
Hone, a five-month-old startup building AI "Engines" that take ownership of business goals for weeks or months, just raised a $60 million seed led by Benchmark and Index.
0:18
Welcome to Startup Bios, I'm Brian McCullough.
0:20
Like and subscribe to hear about the newest startups coming out of stealth and raising new rounds almost every day.
0:26
$60 million is a really big number for a seed round, and Hone is raising it at five months old.
0:32
Benchmark and Index Ventures led the round, and the company says it's now valued at $285 million, according to Bloomberg.
0:41
That valuation figure comes from Hone itself, by the way, and nobody's said whether that's pre or post money, so I'd hold it a little loosely.
0:48
Elad Gil and SV Angel are in as well, and so is Scott Wu, the CEO of Cognition, who invested personally.
0:56
Peter Fenton from Benchmark and Shardul Shah from Index are both joining the board. So what is Hone actually building?
1:07
The CEO, Moritz Stephan put it this way to Bloomberg, quote, we are doing is building AI that owns outcomes."
1:14
End quote. That's kind of the entire pitch.
1:17
We've been talking a lot lately about AI agents, and most of what's out there is agents that do tasks.
1:23
You give it a job, it does the job, and then it waits on you for the next one. Hone wants to go a level above that.
1:30
You don't hand their system a task.
1:32
You hand it a goal, and the goal might take weeks or even months to hit.
1:37
Hone calls these things Engines.
1:39
The example they use on their launch blog is something like "improve customer retention." You assign an Engine to that goal, and then, the way they describe it, the Engine onboards itself.
1:51
It learns the company's systems, the people, the history and the constraints, so it can figure out what good actually looks like there.
1:59
Then it decides on its own what needs doing to hit that goal and who it needs to pull in help, inside whatever guardrails you've set.
2:07
And because it owns a goal instead of a task, Hone says an Engine can come back and propose new things it could take on, or new ways of going about the work. Bloomberg's example was sales.
2:20
Today you can buy software that helps a sales team figure out which incoming leads look promising.
2:26
Hone's idea is that instead of a tool that scores the leads, you'd have an Engine whose job is getting more of those leads to actually convert, and it just keeps working on that, proactively, over time.
2:38
Stephan told Bloomberg the technology is meant to, quote, "look at what a team does, its goals and responsibilities, and deploy engines to own these." End quote. And it checks in with a human when it needs Hone laid out the bigger philosophy in a kind of manifesto on its website.
2:56
AI models keep getting smarter, but the world doesn't really look that different than it did five years ago, and Hone thinks the reason is that today's agents are still basically tools inside workflows that humans run.
3:09
A person has to break the goal down into tasks, feed the agent context, and catch it when it screws up.
3:15
So no matter how good the model gets, you're limited by how much human attention there is to go around.
3:20
Their analogy for this is electricity.
3:24
Hone points out that something like three decades went by between Edison's dynamo and Ford's assembly line, and the big productivity gains didn't come from making more power, they came from reorganizing the factory around it. So Hone's line is, quote, "The dynamo is here.
3:41
At Hone, we're building the factory."
3:43
End quote. Instead of telling the AI what to do step by step, you tell it the end state you want and what not to break.
3:52
Their example is something like, cut drop-off in your onboarding funnel to 8%, % and keep the other metrics steady while you're at it.
3:59
they actually describe it as quote.
4:01
less chatbot and more Kubernetes cluster.
4:04
If you're not a developer, Kubernetes is software that keeps servers in the state you asked for and fixes things on its own when they drift.
4:12
Hone also wants its Engines to stick around.
4:15
Most AI right now, the session ends and whatever it learned kind of evaporates. Hone says Engines are full-time and permanent, and they're supposed to get better the longer they're in a role.
4:26
And the company's pitch is that all of that accumulated knowledge belongs to you, the customer, as your IP, not to some model everybody else is also using. And how you grade one of these things?
4:38
Hone says the real test is the profit and loss.
4:41
It's not about how slick a demo is or how many tasks it knocked out.
4:47
The way they put it in the blog, you judge an Engine the same way you'd judge any serious employee, by asking, quote, "What did you get done this week?" Now, the obvious worry is that you're giving an AI access to sensitive company systems and letting it run for months with a lot of leeway. Hone's answer is guardrails.
5:05
Before it takes on any responsibility, an Engine's decisions get simulated first, anything sensitive needs a human to approve it, and everything gets logged so somebody can go back and look.
5:17
Hone's second blog post goes deep on that simulation piece.
5:21
It opens with Apollo 13, where engineers on the ground tested their fix in a simulator before they sent instructions up to the astronauts.
5:29
Hone's point is that today's AI agents mostly don't do that.
5:32
They make a change and find out in production whether it broke something.
5:37
The thing is, Hone's Engines change themselves.
5:40
They rewrite their own tools, their code, their memories, all as they learn. Any one of those changes can break something that used to work.
5:49
So before an Engine locks a change, it replays it against past situations to see what else it might mess up.
5:56
Their example goes like this. Imagine a manager named George tells the Engine, check with me before you order from a new supplier.
6:04
So the Engine writes itself a rule, always ask George first.
6:08
But when they replay that rule against past situations, it turns out it breaks something. There was a case where a production line was down and somebody said order the bearings now, and with the new rule, the Engine would've sat there waiting on George instead.
6:22
So the rule gets fixed to, always ask George, unless the line is down, and it only goes live once it passes the old cases too.
6:31
The hard part, Hone says, is that you can't rewind a company.
6:35
You can't roll your sales force back two weeks.
6:38
So they built an agent they call the "director" that plays the role of the outside world during a replay, deciding when a person writes back or when a shipment notice shows up. Other agents stand in for each software tool, and Hone says those fake tools got the real facts right about 90% % of the time in tests. They also fast-forward the clock, so weeks of work can replay in minutes. Hone claims that across 125 test scenarios, replaying a change cut the share of changes that broke something from 33 % down to 5%. That is their own internal test, of course, and nobody outside has checked that yet. The founder, Stefan, is Austrian, and he's been starting things since high school.
7:21
Back in Vienna, he and a partner launched a startup called Lobu that delivered books from a small local bookshops by bike as an alternative to Amazon, and they even pitched it on Austrian TV.
7:32
As a teenager he also co-founded Tumbleweed, a project to build a wind-powered rover for Mars that came out of a European Space Agency incubator program. It grew to more than 50 people across four countries, and prototypes got tested in desert simulations of Mars missions in Oman and Israel. Then he went to Stanford for computer science and did AI research in Chelsea Finn's lab.
7:54
and that work got accepted at NER IPS in 2022 while he was there he also started an anonymous campus app called Fizz with Carlo Kobe and Fizz went through Y Combinator After that, Stephan joined Cognition, the company behind AI coding agent Devin, as one of its first employees. He worked on how Devin plans and carries out its work, and eventually became chief of staff to Scott Wu.
8:18
Kobe is one of Hone's cofounders too.
8:22
He's a Thiel Fellow who left Harvard to start Fizz and ran it for several years. The third cofounder, Oliver Brady, was one of the early engineers at the legal AI startup Harvey and then helped build the enterprise business at Mercor. As for customers, right now we know of two, Cognition and the AI inference company Modal.
8:40
With Cognition, Hone is building agents to help with the go-to-market side as Cognition grows. And Shardul Shah from Index basically made the investment case in one sentence to Bloomberg, quote, "What Cognition is to software development, Hone can be to every other function in an organization." End quote. the First thing I'd point out is the Cognition connection. Cognition is an early customer, Cognition's CEO is a personal investor, and Hone's CEO was his chief of staff.
9:08
That's how a lot of startups land their first customer, so it's not a knock. But it does mean the one customer we know the most about is the one that knows the founder best. We haven't seen revenue, we haven't seen results from a customer, and the valuation number is the company's own.
9:22
So for now this is a bet on a team and an idea.
9:25
and the idea is a really hard one.
9:28
A coding task is pretty bounded.
9:31
The tests either pass or they don't.
9:33
A business goal like "improve retention" stays open for months while the market and everything else moves around was a research paper this past June from CEO Bench that had AI models run a fake startup for 500 simulated days, doing things like pricing and marketing and budgeting, and every model they tested did worse than a simple rule-based baseline.
9:54
That's a simulation, of course, and it says nothing about Hone directly.
9:57
But it gives you a sense of how hard long-horizon business decisions still are for these models right now.
10:03
The other problem with owning an outcome is figuring out who gets the credit when a number moves, the Engine or just a good quarter.
10:12
Hone even lists grading and credit assignment over month -long stretches as one of the problems it's working on, so they clearly know that one's coming. The competition is pretty much everyone.
10:23
Decogon and Sierra are well-funded and already selling AI agents to businesses, Salesforce has Agentforce, and OpenAI keeps pushing deeper into agents for office work. Hone's argument would be that those are mostly task tools, and it's going after the layer above them.
10:39
Peter Fenton brought up security, and I think he's right to.
10:43
He told Bloomberg that Hone is working on guardrails to protect customers if one of its agents causes a security incident, which is a fair worry after a run of cases where AI tools went rogue, and he said new technology needs, quote, "appropriate immune systems."
10:58
An agent that runs for months with access to your systems is much bigger as a target than a chatbot. The replay testing idea looks like a real engineering answer to part of that problem, at least on paper.
11:11
though, I'd admit. So yeah, $60 million is a lot of rope for a five-month-old company. If Hone can get from AI-native early adopters like Cognition and Modal to boring, old-school big companies, and actually show them a number moving on the P &L line, this round will probably look cheap. If it can't, well, well, that's a really expensive way to learn that companies still want a person owning the outcome, at least for now.
11:35
That's it for this episode. Like and subscribe.
11:37
If you'd like to get your founding team interviewed on this podcast for 45 minutes or so, get in touch at hi at startupbios .tech.
11:47
That's hi at startupbios .tech and we'll schedule something.
11:51
And if you'd like to get this in your ears every morning, check out the podcast, Startup Bios, wherever you get your podcasts.