Hone raises $60m seed from Benchmark and Index for AI engines

Hone raises $60m seed from Benchmark and Index for AI engines

8 October 2026: Hone, a San Francisco start-up building AI that is assigned business outcomes rather than tasks, has raised a $60 million seed round from Benchmark and Index Ventures. The company announced the money in the same post that launched its first product, which it calls Engines.

An engine is given a goal, not a prompt

Hone's launch post defines the product against two things buyers already know, the chatbot and the agent. An Engine, the company says, "meets goals rather than completing tasks". An operator assigns it an outcome, and the example Hone gives is improving customer retention. The Engine then onboards itself: it learns the company's systems, people, history and constraints so that it knows what a good result looks like in that organisation. It decides what to do and whom to involve, within the guardrails it is given, and it carries on working after any single conversation has ended.

Hone lists four properties. Engines are persistent, so they accumulate context over time instead of starting each session cold. They are team players, pulling in colleagues and learning from feedback. They have KPIs, and the company's own standard for success is the profit and loss account: "Our ultimate eval is the P&L", the post says, adding that customers should judge an Engine by asking what it got done this week. And they have what Hone calls motion, meaning they can spot problems and opportunities without waiting to be told.

One claim in the post matters for anyone weighing data-protection terms. Hone says the operating knowledge an Engine builds "permanently becomes part of the IP of your organization" and does not sit inside a model available to everyone else.

Who backs it and who is building it

The round is $60 million, and the company names Benchmark and Index as the investors. The post does not give the use of the money in detail beyond building Engines, and says Hone is now hiring. Hone does not state a valuation in its own announcement, and we do not repeat figures that appear only in press coverage.

The team describes itself as alumni of Cognition, Decagon, Ramp, Mercor and Jane Street. An earlier post from 12 August, written as a mission statement, said the founders were early contributors to Cognition, Mercor and Ramp and that they work with frontier AI companies and large enterprises on hard problems. That post lists the research areas the team is working on: tools that build, orchestrate and improve agents in production, system components reliable enough to run autonomous organisations, grading and credit assignment across month-long runs, interfaces for people working with agents, and self-constructing ontologies.

The posts are written by Moritz Stephan, who authored the launch announcement, and by Carlo Kobe. A companion research note published on 8 October, the first in a series on simulations for self-improving agents, runs to 15 minutes of reading on Hone's blog.

What a buyer can and cannot learn yet

The product is described entirely in Hone's own words. The site carries endorsements from named customer contacts at companies including Modal, Lucanet and Cognition, but the company has not published pricing, deployment numbers or independent results for Engines. The claim that an Engine can own an outcome such as retention for weeks and months is a design goal in the post, not a measured result.

For readers who evaluate agent platforms, the useful distinction is the unit of work. Hone proposes that its product be judged per outcome, which would require vendor and customer to agree on a metric, a baseline and the systems the Engine may read and change before it starts.

The company's mission statement is dated 12 August and the seed was announced two months later. The company's website lists San Francisco as its base.

Company profile on TrustList: Hone

Sources

• Hone: Introducing Engines: AI with Motion, 8 October 2026
• Hone: Hone builds AI that owns outcomes, not tasks, 12 August 2026