On October 8, 2026, a San Francisco startup named Hone announced a $60 million seed round and introduced a product it calls Engines: persistent AI systems assigned to business outcomes rather than individual tasks. Benchmark and Index Ventures led the financing, valuing the company at $285 million. The announcement lands at a moment when most enterprises acknowledge AI has delivered strong returns inside software development and weak returns everywhere else.

“The ROI of AI outside of engineering today is quite disappointing. That’s what we’re here to solve,” said CEO Moritz Stephan in an interview accompanying the launch.

What Is an Engine, and Why Does It Matter?

Hone’s framing is deliberately different from the “agent” and “copilot” vocabulary that has dominated enterprise AI marketing for the past two years.

An agent, in the dominant usage, completes a task when prompted and then stops. A copilot assists a human who is still doing the driving. An Engine, by Hone’s definition, is assigned a goal (“improve customer retention”) and then operates continuously until that goal is achieved, improving along the way.

The four-step workflow Hone describes:

  1. You assign the Engine to an outcome.
  2. It onboards itself, learning the company’s systems, stakeholders, history, and constraints.
  3. It figures out what to do, collaborating with relevant teammates and adapting when circumstances change.
  4. It delivers the outcomes and continues improving as the company changes.

“This is different from a chatbot or an agent,” the launch post reads. “An Engine meets goals rather than completing tasks, it figures out what to do rather than waiting to be prompted, and it stays focused on its goals even when interactions end or circumstances change.”

That last clause is the operative one. Most enterprise AI today is episodic: valuable within a conversation window, invisible outside it. Hone is betting that the next unlock in enterprise AI value requires crossing from episodic to persistent, from task completion to outcome ownership.

The Enterprise AI Productivity Gap

The timing reflects a real problem that enterprise buyers have started to name explicitly. AI-assisted code generation now routinely reduces engineering cycle times. But when the same companies try to apply the same agentic patterns to sales, account management, procurement, or financial risk, they rarely get equivalent leverage.

Several structural reasons account for the gap. Engineering tasks are well-defined, short-horizon, and have clear success signals (the code compiles or it does not). Business functions involve ambiguous goals, weeks-long execution windows, and success measured on a P&L rather than a test suite. Episodic agents built for the former structure fail at the latter.

Hone’s CEO frames the challenge in terms of accountability: “AI with responsibilities should be permanent and full-time, accumulating context and judgment over time without interruption.”

That framing also explains the product’s name. An Engine, unlike an agent, implies something that runs continuously, not something that is switched on to handle a prompt.

Self-Improving Architecture

The technical approach Hone published alongside the launch announcement addresses one of the harder problems in autonomous AI: how to let a system improve without requiring humans to supervise every iteration.

Hone’s answer is a simulation-based self-improvement cycle. A “director” agent replays weeks of company activity in minutes using fast-forwarded time, exposing the Engine to historical decision points before promoting changes to production. Across 125 distinct test scenarios, that approach cut regressions from 33 percent to 5 percent.

The self-improvement loop follows four stages: Discover (identify where the engine is making mistakes), Replay (simulate the context with time compression), Propose (generate a candidate fix), and Promote (gate the fix against the regression suite before deploying it).

The multiplayer design reinforces the self-improvement loop. Because an Engine serves an entire team rather than an individual, a correction from one person immediately improves the system for everyone. Institutional corrections, not just individual corrections, accumulate in the Engine’s context.

“Your operating knowledge doesn’t disappear at the end of a session or live inside a model available to everyone else,” the Hone launch post states. “It permanently becomes part of the IP of your organization.”

The Team: Cognition, Harvey, Ramp, Midjourney

The founding team’s provenance is notable because the institutional knowledge behind agentic AI sits in a small number of companies that have been operating real production systems at scale.

CEO Moritz Stephan joined Cognition (the company behind Devin) as an early employee and owned the planning-and-execution architecture at the core of that product, including most of Cognition’s long-horizon agent work. He later became Cognition’s Chief of Staff, responsible for company strategy, hiring, and product direction. Scott Wu, Cognition’s CEO and Stephan’s former boss, invested personally in Hone.

Co-founder Oliver Brady was an early engineer at Harvey, the AI legal platform, and then led the enterprise effort at Mercor. Co-founder Carlo Kobe is a Thiel Fellow who previously co-founded Fizz with Stephan.

The broader team has pulled engineers and operators from Ramp, Midjourney, Decagon, Jane Street, and several others who set aside plans to start their own companies to join instead.

Index Ventures partner Shardul Shah, who joined Hone’s board alongside Benchmark partner Peter Fenton, summarized the thesis in the firm’s investment announcement: “As frontier models handle ever-longer and more complex tasks, we believe value will accrue to whoever owns the layer between the model and the enterprise.”

Early Customers and Beachhead

Hone’s first disclosed customers are Cognition (the company Stephan came from) and Modal, a fast-growing AI inference startup. Both are using Engines for go-to-market work, which aligns with the company’s stated beachhead: account management and revenue growth.

The account management use case illustrates the product’s design philosophy. A traditional CRM workflow asks a sales rep to review signals, decide what to act on, and then act. A Hone Engine is assigned to a portfolio of accounts, monitors signals continuously across Slack, email, and connected systems, identifies which relationships need attention, and takes action or queues the action for human approval, depending on governance settings.

For more sensitive actions, Hone requires explicit human approval before execution. All actions are logged, and execution runs inside a sandboxed environment. Customer data is not used to train Hone’s underlying models.

Outcome-Based Pricing

How Hone charges for its product is as much a product decision as a business model decision. The company is exploring pricing tied to outcomes or budgets rather than seats or task counts. One model it has discussed publicly: give the Engine a fixed spending budget and let it decide how to allocate that budget across context retrieval, tool calls, and compute to achieve the assigned outcome.

That framing has a direct precedent in the conversational AI layer: Sierra, which handles customer-facing AI for enterprises, charges per resolved outcome rather than per token or per seat. Hone is attempting to apply that logic to internal enterprise functions.

“What did you get done this week?” is how Hone says customers should evaluate their Engines, using the same standard they would apply to a human contributor.

Where This Fits in the Agentic Enterprise Stack

The enterprise AI agent market has been layering for two years. Infrastructure companies (compute, inference, orchestration) built the foundation. Security companies (runtime monitoring, access governance, policy enforcement) are building the guardrails. Vertical companies (legal, finance, healthcare-specific agents) are capturing domain value.

Hone is positioning itself in a fourth layer: the organizational ownership layer, sitting between a company’s systems of record and the frontier models that power agent reasoning. Its thesis is that long-term enterprise value flows to whoever makes the organizational knowledge of a company durable, compounding, and transferable to AI systems without losing it at session end.

That is a different value proposition than productivity at the task level. It is closer to the value proposition of enterprise software broadly: persistent state that survives personnel turnover and compounds with use.

For enterprise AI buyers evaluating their agent roadmap, Hone’s launch is worth tracking for two reasons. First, the team’s background in real production agent systems (especially at Cognition) suggests the technical claims are grounded. Second, the outcome-based framing clarifies an accountability gap that has kept enterprise AI adoption episodic: the absence of AI systems that can be held to the same “what did you get done this week” standard as a human contributor.

The $60 million seed at a $285 million valuation is a strong signal that top-tier investors believe that gap is now closeable.

Sources:

For more on how enterprise teams are redesigning workflows around AI systems that own outcomes rather than complete tasks, see Enera’s analysis of autonomous GTM systems and the enterprise AI adoption readiness gap. For examples of the agentic stack these systems run on, see our coverage of Manus’s $500M autonomous agent round.