Hone AI Agents Startup Raises $60M Seed to Own Business Outcomes

A seed round sized like a Series B

Hone is a five-month-old startup founded by people who previously worked at OpenAI and Cognition AI. It has raised a $60 million seed round to build AI agents that act less like software tools and more like staff members who take on assignments lasting weeks or months1. Benchmark and Index Ventures led the round, and the company says it is now valued at $285 million1. Elad Gil and SV Angel also invested, and the post-money figure implies a pre-money valuation of $225 million2.

CEO Moritz Stephan was previously chief of staff at Cognition, the company behind the AI coding agent Devin1. He also has a research background at Stanford2. Other members of the founding team come from OpenAI, Mercor, Ramp and Jane Street2. Stephan's description of the company is short: Hone is building AI that "owns outcomes." The system studies a team's goals and responsibilities, deploys software to carry them out, and checks in with a human when it needs to1.

The outlets that covered the raise agree on the basic facts: the amount, the valuation, the lead investors and the early customers. They differ on how much the pitch has been proven. Bloomberg called it a new entrant in a crowded field1. CryptoBriefing read the valuation as a sign that investors expect a large market rather than a niche product2. RuntimeWire was the most cautious. It noted that the reporting includes no revenue figures, no customer results and no independent performance data15. In my view, RuntimeWire has the most useful framing. For now, Hone is a bet on its team and its distribution channels, not on results anyone has measured.

The product idea: from point tools to "engines"

Hone calls its persistent agents "Engines." The concept is easiest to see by comparing it with existing sales software. A typical AI sales tool might score incoming leads. Hone says its engine would take on the job of raising lead conversion over time, working on its own initiative rather than waiting for instructions115. The company has also named procurement savings as an example. In both sales and procurement, results build up slowly, so the system has to track its own progress and change course as it goes2.

According to Hone's own description, a customer assigns an engine an outcome, connects it to company systems and sets limits on what it can do. Before an engine is given more responsibility, its decisions are simulated. Sensitive actions require human approval, and every action is logged for review15. RuntimeWire stresses that these are Hone's claims about its own system and have not been independently evaluated15. Summaries of the launch also present Hone as a replacement for legacy software rather than something layered on top of it16.

The company describes its mission as closing the gap between what frontier models can do in principle and what they actually deliver inside companies2. That framing matters for the business model. Hone does not plan to train its own frontier model. It is betting that the money lies in the deployment layer: knowing how to break a company's goals into long-running work an agent can manage, and making that work safe enough for executives to hand over.

Go-to-market: start inside the family

Hone's choice of first customers is the most revealing part of its strategy. Cognition and the AI inference startup Modal are both early customers1. The Cognition work focuses on building agents for Cognition's own go-to-market operations as that company grows quickly1. Cognition CEO Scott Wu is also a personal investor in Hone1.

RuntimeWire spelled out what that means. Hone's best-known customer is the former employer of its CEO, and that customer's chief executive is also an investor15. That does not discredit the deal. Selling first to people you have worked with is one of the most common founder tactics. It does mean Cognition is a warm design partner, not proof that strangers will buy. CryptoBriefing raised the obvious next question: can Hone move beyond AI-native early adopters to traditional enterprises2?

There is a sound reason to start with go-to-market work at a company growing as fast as Cognition. Cognition recently raised more than $2 billion at a $48 billion valuation, and its annualized revenue run rate grew from $492 million to nearly $900 million between its May Series D and its September Series E17. A company scaling that fast has a lot of sales and operations work to absorb. That makes it a demanding test of whether agents can take on functions that would otherwise require more hiring. If Hone's engines can visibly help Cognition's sales effort, it gets a strong reference customer for the rest of the market.

The investor thesis: "Cognition for every other function"

Index partner Shardul Shah, who is joining Hone's board, summed up the thesis in one comparison: Hone could be to every other business function what Cognition is to software development1. He also said industry-specific point solutions will still create a lot of value1. That is a fair concession, and it shows the real competitive question. Hone is betting that one general-purpose agent platform can beat specialists in many departments at once.

The comparison also explains the price. Cognition's two latest rounds valued it at about 53 times run-rate revenue each time, according to an analysis cited in coverage of its Series E17. Investors who watched a coding-agent company reach that scale have an obvious reason to pay seed-stage prices for a team that promises the same thing for sales, procurement and operations. The founders' Cognition background is not incidental to the deal. It is the deal.

Benchmark has backed similar companies before. It led a round for Applied Compute, a startup founded by three former OpenAI staffers that helps enterprises train and deploy custom agents, at a $100 million valuation8. Hone's $285 million seed price shows how quickly expectations for this category have risen.

A crowded field with different philosophies

Bloomberg was right to call the market crowded1. In the same week Hone announced its round, Nous Research said it had raised a $90 million Series B at a $1.5 billion valuation and launched Hermes for Businesses, an enterprise agent product built on open-weight models18. That coverage listed Sierra, Cognition, Decagon, Glean and Harvey as well-funded companies with closed agent platforms that all pitch some version of the same idea: let an agent run a multi-step business process18.

Other former OpenAI staff are working on the same problem from different directions. Angela Jiang, an early OpenAI product manager, launched Worktrace AI, which observes how employees work and automates their workflows. Its backers include Mira Murati and the OpenAI Startup Fund69. Andon Labs, which runs real AI-operated businesses as safety testbeds, launched Pion in September 2026 as a research preview of a platform for agents that run whole businesses. The frontier labs are also moving into deployment. In May 2026, Anthropic worked with Blackstone, Hellman & Friedman, Goldman Sachs and others to create Ode, a services company focused on integrating AI into businesses13.

These companies are taking several different approaches. Some, like Nous, compete on openness. Some, like Worktrace, start by observing how people work. Some, like Ode, are services businesses backed by large financial institutions. Hone's approach is ownership of outcomes. It sells a result, not a tool or a consulting engagement. If that holds up, it points toward outcome-linked pricing rather than per-seat software licenses, although no coverage has reported how Hone actually charges.

Security is the real constraint

Benchmark's Peter Fenton, who is also joining the board, said Hone is building safeguards against cybersecurity incidents involving its agents. He framed this in light of recent breaches involving AI tools going rogue1. CryptoBriefing explained why this matters especially for Hone. An agent that runs procurement or sales needs access to sensitive systems and data11.

I think this is the central tension in the company. The more a system "owns" an outcome over months, the more access and autonomy it needs, and the more damage it can do if something goes wrong. Hone's planned safeguards are simulated decisions, human approval for sensitive actions and full logging15. They are the right design choices, but they also slow the agent down. Enterprise buyers will judge Hone less on how autonomous its engines are and more on how well those limits work in practice.

The reading

Hone is a premium bet on a team with a credible pedigree, an obvious first customer and a thesis that follows from Cognition's success. All of the coverage confirms the money and the ambition. None of it confirms traction outside the founders' own network. In the next year, Hone needs to show that its engines raise measurable numbers, such as conversion rates or procurement savings, at a company where no one on the board has a personal connection. Until then, the $285 million valuation reflects how much investors want an agent company that can do for business operations what Cognition did for coding, not evidence that Hone can.