AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
At AI Engineer New York, Tenex cofounder Alex Lieberman returned to the Latent Space podcast with CJ Hess, the firm's director of engineering and its first hire, to describe what happens when a pay-for-output model is scaled into a real engineering organization. Tenex positions itself as an applied AI partner that helps large companies move from what Lieberman calls AI absent to AI native. The company started about 18 months ago as an AI-powered development shop and now reports 75 people, about 90% of them engineers or AI strategists, the latter functioning as technical consultants.
Customer mix has shifted upmarket. Lieberman said roughly half of Tenex's customers are Fortune 2000 companies and half are mid-market businesses doing $100 million to $500 million in revenue. Early customers wanted staff augmentation and AI engineering help; increasingly they ask for AI transformation across the whole company. That reframing matters because it changes the shape of the work from shipping features to changing how an enterprise operates.
On compensation, the story points model survives as an important input. The standard objections to paying by output are that people will inflate estimates or ship sloppy code to rack up points. Lieberman says neither has been the main problem, which he attributes to a deliberate incentive mismatch: AI strategists are measured on customer retention and growth, so they have reason to reject work that does not serve the client. The real difficulty is compensating time that never becomes a ticket, such as architecting, sitting with a customer, or deciding what should be built. Hess said he codes less as models improve and spends more time getting a customer's context into a form a model can consume, with models now reaching roughly 90% of a solution when given the right context. His expectation is that evaluation will drift toward the quality and shape of the output while still anchoring to story points as the unit.
Lieberman described a roughly 4x spread between the highest and lowest paid engineers at Tenex, a narrower dispersion than the 10x some listeners expected from earlier conversations about million-dollar engineers. He frames the model as a leveraged bet on a person's intellect and work ethic, and concedes it is not for everyone. Hess's own path illustrates the logic: as an hourly consultant he watched his output rise while his pay stayed flat, and joined partly on the strength of Lieberman's public pitch about rewarding output over time.
The most concrete engineering detail is MetaHarness, Tenex's internal workflow layer. Hess described it as a standardized set of skills that grew into structured documentation inside each repository, wrapped by a CLI that models use to drive work: drafting plans, producing visualizations, generating client readouts, and generally automating the pieces around writing code. A canonical plan document has a common structure, though Hess said the team has loosened the enforced structure as models have improved. Agents log their decisions and activity in Markdown and increasingly HTML, and all of it is checked into Git alongside the code.
The design rationale is that code alone records what was built but not why. By keeping plans, acceptance criteria, decisions and evidence in the repo, an agent returning to a task can recover the reasoning, and nothing important lives only in someone's head or in a chat thread. Tenex's published article on its AI-native SDLC presents this as a project record that humans and agents both work against, with views into Linear, GitHub and client updates. The article also flags the cost: that much process can be overkill for a small team on a single project. This is the trade-off worth noting for teams considering something similar. Structure buys legibility and continuity across long-running agent tasks, but it is overhead that has to be maintained, and the payoff depends on scale and turnover.
Tenex measures whether the process works by tracking average story points per engineer per month. Lieberman estimated output has roughly doubled this year, pointing to late-December model releases as the inflection for long-running tasks. That figure is a self-reported internal trend, not an independent benchmark, and story points are only as meaningful as the quarterly recalibration behind them. Hess acknowledged the team revisits its estimation criteria at least once a quarter; in principle a point should deflate as models get cheaper and faster, which is an awkward thing to manage when points also feed pay.
The conversation also covered Kill My SaaS, a Latent Space experiment in which builders were asked to replace a speaker-management tool quoted at $43,000 a year. It drew 72 submissions. The team that won with human judges placed 19th on an AI judge's leaderboard, a reminder that checking software against a list of scenarios and judging whether people actually want to use it are different evaluations. Lieberman's pushback on build-versus-buy: ripping out an established system gets more expensive and riskier further upmarket, and any replacement needs someone to maintain it.
On vertical AI, Lieberman is broadly skeptical, questioning how one-off enterprise solutions can be delivered at software margins and how much of a task frontier models will absorb. The counterargument cites Harvey, OpenEvidence and Bloomberg: domain data, specialized interfaces, integrations and habits can hold value even when the underlying model is a commodity. Hess separates vertical products whose main feature is an agent doing a task from tools that improve the workflow around it.
Why it matters: teams building agent-assisted development can take a concrete pattern from this, namely checking plans, decisions and rationale into the repository so agents inherit context instead of rediscovering it. Product leaders weighing build-versus-buy should note that the maintenance burden Lieberman raises is a recurring cost, not a one-time build price, and that cheaper generation does not by itself settle what work is worth. My inference is that output-based compensation will stay rare outside firms with unusually tight customer feedback loops, because the non-coding work Hess describes is exactly where point systems strain.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Listen now (44 mins) | Paying for outcomes, building with agents, and the last mile of enterprise AI adoption