10x leaders explain how they organize AI engineering—and pay for output

At 10x, coding agents get more than source code: they get plans, decision logs and the context behind the work. In a Latent Space interview published October 21, co-founder Alex Lieberman and special projects lead CJ Hess explained how that shared record supports an engineering business that largely pays for output rather than time.

The repository holds the reasoning

Hess described a workflow built around structured documents in a code repository—the shared project store—and a command-line tool that models use. The system supports planning, visualizations and client readouts, not just writing software. Its purpose is to make the surrounding work more automatic and consistent across the team.

1. Start with a central plan. The team maintains a main planning document with a common structure, giving agents a record of what they are meant to build.
2. Record decisions during implementation. Agents follow a defined approach to logging their choices and activity, rather than leaving that history outside the project.
3. Save the context alongside the code. Plans, logs and related material are checked into Git, keeping them available for later questions from engineers and agents.

The distinction is between knowing what the software does and knowing why someone chose to build it that way. Hess argued that code alone cannot replace the history of decisions and customer conversations. Keeping that history in the repository gives the team a durable record instead of relying on someone’s memory.

This does not mean imposing ever more rigid paperwork. Hess said the team has relaxed some document-structure requirements as models have improved. His stated goal is to concentrate context in one place so agents can drive more work without people repeatedly supplying the background.

Faster work should not mean lower pay

Hess’s interest in the compensation model came from his own consulting experience. As AI helped him deliver better work faster, hourly billing meant earning less. He joined 10x attracted by the idea that engineers could benefit financially from higher output rather than be rewarded for taking more time.

At 10x, compensation remains largely tied to output, anchored in story points: agreed scores assigned to engineering work. Lieberman estimated that the highest-paid engineer earns roughly four times what the lowest-paid engineer does. He described that variation as part of the model, while acknowledging it is not for everyone.

The obvious objections came up in the interview: engineers might inflate subjective scores or rush out poor code. Lieberman said those had not become problems, crediting AI strategists whose incentives depend on customer retention and growth. That is his assessment of the company’s controls, not an independent finding about code quality.

The work that coding scores miss

The bigger compensation difficulty, Lieberman said, has been accounting for time spent on architecture and client meetings. The company has had to add nuance to its pay model. Hess sees growing value in listening to customers and translating what they need into context a model can use.

That work serves a business Lieberman said now has 75 employees, about 90% engineers or AI strategists. Half its customers are Fortune 2000 companies; the rest are mid-market businesses with $100 million to $500 million in revenue. He said customers increasingly wanted broader help adopting AI, not simply extra engineering staff.

Hess suggested future evaluation could focus more on how finished work functions and looks, while retaining story points as an anchor. That was a direction he envisioned, not a replacement compensation system announced in the interview.