Teamday
← Back

Tenex: Alex Lieberman and CJ Hess on AI-Native Engineering

Latent Space
agentsenterprisefuture-of-workknowledge-workbusinessautomationinterview

How Tenex Builds an AI-Native Engineering Team in New York

Tenex is an applied AI partner that helps large enterprises become AI-native. Co-founder Alex Lieberman, who previously co-founded Morning Brew, and CJ Hess, Tenex's first hire and now its special projects lead, sat down with Latent Space at the AI Engineer conference. Tenex started 18 months ago as an AI-powered dev shop. It now has 75 people, 90% of them engineers or "AI strategists" (technical consultants), and half its customers are Fortune 2000 companies. The other half are mid-market firms with $100 to $500 million in revenue.

Paying engineers by story points, not hours: Hess joined after running his own engineering consultancy, where AI made his output faster while hourly billing paid him less. Tenex instead pays engineers per story point delivered. Lieberman frames it this way: "AI is really just a leveraged bet on a human's intellect and work ethic." The gap between Tenex's highest- and lowest-paid engineers is about 4x. The feared failure modes, sandbagged estimates and sloppy code, haven't appeared, because AI strategists are paid on customer retention and growth, which pulls against engineers. The harder problem has been paying for the work that isn't coding, such as architecture and client meetings.

The Tenex meta harness: Hess describes the "meta harness": a standard set of skills, a structured documentation layout in each repo, and a CLI the models use to drive the work, from plans and visualizations to client readouts. Every project has a canonical plan document, and agents log their decisions and progress in a fixed way. Everything is checked into Git, so the record of why something was built sits next to the code: "It's never like in someone's head. It's always in the repo." Tenex measures whether this works by tracking average story points per engineer per month. Hess estimates output has roughly doubled this year, starting with the Opus 4.5 "moment" last December. The team recalibrates what a story point is worth at least once a quarter.

Tools and token spend: Most of the team defaults to Claude Opus 5.5 and uses Codex to review its work, since having the two go back and forth still gives the best results. Hess works in Amp's cloud "orbs", and says plugging in a Claude Code subscription cut costs sharply. Asked about token spend against salary spend, the team puts Tenex somewhere between 1:1 and 2:1. Lieberman thinks cost per task has fallen far enough that one engineer token-maxing is much less of a threat to the company than it was six to eight months ago.

The enterprise AI adoption ladder: Lieberman's numbers from about 100 enterprises Tenex talks to: 90% have rolled out Claude or Codex to employees, about 60% have done real training and enablement, and only about half have rolled out coding agents to their engineering and product teams. The biggest drop comes at the next step: "agents in production for non-engineering work drops to like five percent." The most common requests are back-office agents for repeatable work and data permissioning (RBAC). Every employee now has an assistant that needs access to data, which Lieberman calls a nightmare for CIOs and CISOs. Tenex first built permissioning gateways for each client and has since turned it into its own gateway product.

Standardize first, then automate: Hess's rule for agent projects is "you can't automate a process that does not exist." When 40 people do the same job 40 slightly different ways, a background agent has no definition of "right" and there's nothing to write evals against. Tenex often has clients agree on one process first, then hand it off to an agent, so employees move to reviewing work the agent surfaces.

5 Takeaways from Tenex on Running an AI-Native Services Firm

  • Pay for output, with a counterweight - Story-point pay works at Tenex because AI strategists are paid on retention and growth, which keeps engineers from gaming estimates.
  • Put all context in the repo - Plans, decision logs and agent progress live in Git next to the code, so agents and people can see why something was built.
  • Recalibrate the unit - As models improve, a story point covers more work. Tenex reviews its estimation criteria every quarter.
  • Expect the gap in non-engineering agents - Chat rollouts are nearly universal, but production agents outside engineering are rare. Permissioning is a big reason why.
  • Skepticism on vertical AI - Lieberman is bearish on vertical AI apps such as Harvey or OpenEvidence, because buyers want one-of-one software and frontier models keep improving. The host pushes back that those companies are winning on focus and domain-specific interfaces.

What Tenex's Model Means for AI-Powered Organizations

Tenex shows what changes when engineers are paid for output and context is written down instead of kept in people's heads: output per person roughly doubled in a year, and the firm sells the same practices to enterprises. Its adoption numbers are the warning. Most large companies have handed out chat assistants, but few have agents doing real non-engineering work, and the blockers are process standardization and data permissioning more than model capability. On jobs, Hess argues that agents mostly lead to reallocating people toward bigger problems, while Lieberman says employees also need to take the initiative to reskill rather than wait for it to happen to them.