Intel / article

AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess

Source: latent.space
Date: Published Oct 21, 2026

Context

In this Latent Space episode, recorded at AI Engineer New York and published with an article and an automatically generated transcript, Tenex's "AI-native SDLC" turns out to be mainly about where a project's lives. CJ Hess describes MetaHarness, Tenex's shared set of skills, structured project documents and a command-line tool that the models use to drive the work: a canonical plan, a log of the decisions agents make, and readouts for clients, all checked into Git beside the code. The point, he says, is that an returning to a task can read not just what was built but why. Tenex's own write-up, Building an AI-Native SDLC, calls this treating context as code: each project keeps a repository of markdown with current truth, planned work and history in separate folders, and the tool syncs ticket status from Linear and GitHub into reserved blocks without rewriting the prose around them. The same article says this is overkill for one team on one project and pays for itself as more work runs in parallel.

The system is tied to how Tenex pays people. Its engineers are compensated largely by story points, an agreed estimate of each task's size, rather than by hours. Alex Lieberman says the feared problems, inflated estimates and shipping bad code to collect points, have not been the main issue, because Tenex's AI strategists, who lead client engagements, are rewarded for customer retention and growth and push back on work that doesn't serve the client. The harder problem is rewarding work that does not fit neatly into a ticket, such as understanding a customer's problem or designing an architecture. The guests say they judge whether a process change works by the trend in average story points per engineer per month, estimate that output has roughly doubled this year, and recalibrate what a story point is worth about once a quarter. Lieberman also puts the gap between Tenex's highest- and lowest-paid engineers at about four times. These are the guests' own estimates from conversation, not published data.

The episode's title points at Tenex's customers, which Lieberman says are split evenly between Fortune 2000 companies and mid-market firms with $100 million to $500 million in revenue. Of roughly 100 enterprises Tenex is talking to, the guests estimate about 90% have given employees Claude or Codex, about 60% have trained people to use them, and closer to 50% have rolled out coding agents to engineering. Only about 5% have an agent in production doing work other than building software; in the same breath, the guests say most enterprises they talk to with more than a thousand employees have no such agent.

Key takeaways · AI-distilled

Alex Lieberman says the gap between Tenex's highest-paid and lowest-paid engineer is about 4x under its story-point pay model, and that 90% of its 75 staff are engineers or AI strategists.
CJ Hess estimates Tenex engineers' output has roughly doubled this year as models took on long-running tasks. Tenex reviews its story-point estimation criteria at least once a quarter so pay keeps tracking the value delivered as models improve.
Of about 100 enterprises Tenex talks to, Alex Lieberman estimates 90% have rolled out Claude or Codex, about 60% train staff to use it, close to 50% have coding agents, and about 5% run agents in production for non-engineering work.
CJ Hess says agents are hard to deploy when 40 people do the same job in slightly different ways. Tenex's advice is to standardize the process first, because without that you cannot build or tell whether the agent is doing the job right.
Latent Space's Kill My SaaS contest drew 72 submissions. The entry its human judges picked ranked 19th on its 's leaderboard, a sign that checking software against scenarios is not the same as judging whether people want to use it.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
LLM-as-judge — Using one model to score another's output against a rubric, so quality can be measured at a scale human grading cannot reach.

Read the source www.latent.space