Hone Raises $60M Seed to Build Persistent AI Engines

Giving software responsibility for a business goal forces a company to explain what success means and who can authorize the work. Hone, a San Francisco applied AI company, is building persistent agents called Engines around that relationship. On October 8, 2026, Hone announced a $60M seed round backed by Benchmark and Index Ventures to develop the systems.

Benchmark and Index led the financing, and Bloomberg reported a $285M valuation according to Hone. The company's official LinkedIn announcement names Elad Gil, Hanabi, Definition, Diffusion, Lux, SV Angel and Align as additional participants. The round finances a product intended to hold an ongoing role inside an organization, retaining its context and adapting its work as circumstances change.

Co-founders Moritz Stephan, Oliver Brady and Carlo Kobe are bringing experience in AI coding, enterprise engineering and company building into that proposition. Hone describes Engines that learn from a team's corrections and test their own changes against earlier work. For an enterprise buyer, that makes the implementation a continuing relationship between the people defining a goal and the software expected to keep pursuing it.

The founders are carrying operating experience into enterprise AI

Index Ventures' investment account provides the clearest public explanation of the founder team. Moritz Stephan worked on Devin's planning and execution harness at Cognition before becoming the company's chief of staff. Oliver Brady was an early engineer at Harvey and later helped build Mercor's enterprise effort. Carlo Kobe co-founded Fizz with Moritz Stephan in 2021 and continued leading the business while Moritz Stephan returned to Stanford to complete his degree.

Those backgrounds connect technical execution with the organizational work around it. An agent's plan has to meet the business where the plan will operate: its permissions, working habits, people and history. The founders' experience gives them several routes into that problem, from engineering an agent that carries out extended assignments to helping an enterprise organization adopt new software.

Index's investment thesis emphasizes the layer between frontier models and the enterprise using them. Hone is positioning its research and product in that layer, where improvements in model capability still have to become useful behavior inside a particular company. The financing backs that development effort; it also brings investors into a business whose customer relationships will depend on understanding how work is already done before asking software to change it.

A shared Engine makes teaching software an organizational task

Hone describes Engines as agents assigned continuing outcomes, with customer retention and vendor management among the examples. The company says an Engine learns the relevant systems, people, history and constraints, collaborates with the team, and carries its knowledge forward. Its product description places that collaboration in Slack, Microsoft Teams and email, where a correction from one person is intended to improve the Engine for other users.

That shared learning changes the implementation discussion. If an employee corrects how a customer commitment should be handled, the correction may affect future work involving several departments. The business needs a way to distinguish a useful lesson from a one-off exception, and to make clear whose instruction carries authority when colleagues disagree. An organization's knowledge becomes more valuable when it travels, but traveling knowledge also needs a responsible owner.

Hone says sensitive actions can require approval and that actions are logged. These controls belong close to the commercial promise because an ongoing assignment can cross boundaries that a single task never reaches. A buyer evaluating an Engine would need to connect its goal with permitted actions and escalation decisions, then keep those choices current as the team's responsibilities evolve. The product therefore asks something of the customer alongside the software: an operational definition of what the Engine is allowed to own.

Self-improvement needs a record of what already worked

Hone's engineering team describes a replay system for evaluating changes before adopting them. Engines can alter their tools, code, memories and other working context, so an improvement can change behavior that previously worked. The published process finds issues in traces, recreates the work, proposes a change, and checks the change for regressions before promotion.

The company reports that, across 125 distinct scenarios, replay before promotion reduced regressions from 33% to 5%. This is an internal engineering result, with a narrower meaning than a measured improvement in a customer's revenue or operating margin. It supports discussion of the testing approach; the launch materials do not provide independently measured financial returns, deployment costs or revenue figures.

The underlying engineering difficulty is unusually relevant to the product's ambition. A company keeps changing while software learns how to work inside it. Someone answers a message, a supplier changes a commitment, or a process acquires an exception that was never formally documented. Hone describes simulating people, events and tools to evaluate alternative behavior against that moving environment. For a buyer, the useful question is whether the evaluation record includes the obligations the team would be least willing to lose when an Engine tries a new approach.

The seed finances the work of making delegation durable

Hone's careers page lists roles across agent deployments, infrastructure, product engineering, research engineering, design and business operations. That mix gives the funding announcement a practical shape: the company is recruiting for both the systems behind Engines and the work of bringing them into organizations. It has not disclosed a numerical hiring commitment or a precise allocation of the $60M among those functions.

The surrounding enterprise AI market also makes control part of the buying conversation. DevCuration's coverage of Serval's acquisition of Ensignia and Rein Security's agent security financing follows adjacent efforts to address security around automated work. These records provide context for Hone's governance choices without establishing a partnership, product equivalence or competitive ranking.

Hone is asking customers to make a goal persistent enough for software to carry it across conversations and changes in company state. For an operator evaluating that proposition, a performance review would need to connect the Engine's actions with the business outcome, including the decisions that still required human approval. The next budget discussion starts with the goal the team chose and the authority it was willing to hand over.