AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 23, 2026

Alex Lieberman and CJ Hess of Tenex used their appearance at AIE New York to describe how the company scaled its output-based engineering pay model into a 75-person organization, and to walk through the internal workflow system they call MetaHarness. Lieberman founded Tenex roughly 18 months ago with Armand as an AI-powered development shop; Hess was the first hire and now carries the title of special projects lead. The session revisited a proposition the pair first aired on the podcast a year earlier: engineers are paid for the work they deliver rather than the hours they log, on the theory that AI raises output and an hourly rate would therefore punish the fastest engineers.

Tenex describes itself as an applied AI partner that helps large companies move from what Lieberman calls AI absent to AI native. The customer mix has shifted upmarket: about half are Fortune 2000 companies and half are mid-market businesses with revenue between $100 million and $500 million. Of the 75 employees, roughly 90% are either engineers or AI strategists, the latter functioning as technical consultants. Lieberman said clients increasingly want AI transformation as a service rather than simply staff augmentation or AI engineering.

Story points remain an important part of engineering compensation. The failure modes outsiders predicted — engineers sandbagging estimates or shipping sloppy code to ring up points — have not been the biggest problem, according to Lieberman. He credits the incentive structure: AI strategists are measured on customer retention and growth, so they have reason to reject work that does not serve the client, which puts them at odds with engineers looking to book points.

The harder issue is everything that does not fit neatly into a ticket: understanding a customer's problem, designing an architecture, deciding what should be built. Hess said his own time has moved from heavy coding toward collecting customer context and shaping it into a form the model can use, and he estimated that given the right context models now get to roughly 90% of a solution. Lieberman acknowledged that compensating non-coding work has required nuance in the model, which remains largely output-driven. He also said the spread between the highest- and lowest-paid engineer at Tenex is around four times, not the tenfold outsiders sometimes assume, and framed AI as a leveraged bet on a person's intellect and work ethic.

On performance, Hess estimated engineer output — measured as average story points per engineer per month — has roughly doubled over the past year. He dated the shift to a December release he described as an Opus 4.5 moment, when models began handling long-running tasks, and pointed to a newer Opus 5.5 as substantially more capable. Tenex recalibrates the story-point rubric at least quarterly. There is a deflationary tension built into this: if models make a given task easier, a fixed point value overpays for it, so the unit has to be re-based as capability rises. Tenex's answer is periodic recalibration rather than a static rate card.

MetaHarness is the mechanism behind that. It began as a standardized set of skills and grew into structured documentation inside the repository plus a command-line interface the models use to drive work: producing a plan, a visualization, a client readout. A canonical plan document anchors the project, though Hess said the enforcement around its structure has loosened as models have improved. Agents are expected to log their decisions and activity in an opinionated format, and all of it is checked into Git alongside the code — increasingly as HTML picture books rather than plain Markdown. Plans and rationale sit with the implementation, so an agent returning to a task can read not only what was built but why.

The argument for this is legibility rather than raw speed. Code alone can record what exists; the decisions, acceptance criteria and evidence behind it otherwise live in conversations or in people's heads. Keeping them in the repository gives agents durable context and gives humans a place to look when a question resurfaces. Tenex's own write-up on its AI-native SDLC frames this as a project record that humans and agents work against, with views into Linear, GitHub and client updates. The company also notes the cost: that much structure can be overkill for a small team on a single project. In practice, the benefit scales with team size, handoff frequency and how long a project runs — a solo developer on one codebase will likely find the ceremony outweighs the gain.

The conversation also covered Kill My SaaS, an experiment in which the hosts sought replacements for a conference speaker-management tool quoted at $43,000 a year. It drew 72 submissions, and the team that won with human judges placed 19th on an AI judge's leaderboard — a reminder, per the hosts, that checking software against a list of scenarios and judging whether people want to use it are distinct tasks. Lieberman's pushback on build-versus-buy: the further upmarket a company goes, the more expensive and risky it is to rip out an established system, and even a small tool needs someone to maintain it. The hosts countered that ownership buys the ability to improve the workflow and iterate in minutes rather than waiting months on a vendor.

On vertical AI, Lieberman was generally bearish, questioning how a one-of-one solution for every enterprise can be delivered at software margins and how much of a task frontier models will absorb. The hosts pointed to Harvey and OpenEvidence as evidence that domain data, specialized interfaces, integrations and habits can matter even when the underlying model is broadly available, with Bloomberg's terminal as an analogy. Hess drew a distinction between a vertical product whose main feature is an agent performing a task and a tool that improves the workflow around it.

Why it matters: Teams experimenting with agent-heavy development should note that Tenex's gains appear to come as much from legible process — plans, decision logs and acceptance criteria in version control — as from model capability, which is inference from the described workflow but a testable one. Product and compensation leaders weighing output-based pay will face the same boundary Tenex hit: ticket-shaped work is easy to measure, while architecture, discovery and context gathering are where models still need humans, and those roles need a place in the rubric. Buyers deciding whether to replace a vendor with an in-house build should price maintenance and migration risk, not just the subscription.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Paying for outcomes, building with agents, and the last mile of enterprise AI adoption