AI in NYC: How Tenex Builds an AI-Native Engineering Team — Alex Lieberman & CJ Hess
Alex Lieberman and CJ Hess of Tenex discuss how an AI-native engineering consultancy organizes its work around outcomes, not hours. They cover story-point compensation, the MetaHarness workflow that keeps plans and decisions in Git, token economics and model choices, the gap between enterprise AI rollouts and production agents, why vertical AI and mid-tier SaaS may or may not survive, and what this means for jobs and New York's AI scene.
Tenex positions itself as an applied AI partner that helps large enterprises move from AI-absent to AI-native. Its core belief is that frontier models are raw intelligence, and the hardest problem is the final mile of diffusing that intelligence through the economy. Alex Lieberman and CJ Hess use the conversation to explain how Tenex’s compensation model, engineering workflow, and client work are changing as agents take on more of the implementation.
Paying for outcomes instead of time
Tenex started about 18 months ago as an AI-powered dev shop, built on the idea that engineering is the canary in the coal mine for AI. CJ, the first hire, describes leaving a string of failed startups and a consulting practice where his output kept rising while his hourly billing kept his pay flat. Engineers are paid by story points, agreed upon in advance. The thesis is that AI is a leveraged bet on a person’s intellect and work ethic, so earning capacity will vary more than ever. The highest-paid engineer at Tenex earns roughly 4x the lowest-paid one, not the 10x they would like.
The classic objections, that story points are subjective and that people will sandbag them or ship sloppy code, have not materialized. Alex credits AI strategists, who are incentivized on customer retention and growth rather than on points. The harder problem is compensating time that is not spent coding, such as architecture and client meetings. The comp model handles this with nuance but still anchors to output. Story points are recalibrated at least once a quarter. Engineering output has roughly doubled this year, measured from the Opus 4.5 moment in December when models started handling long-running tasks. Whether a point is worth the same as it was a year ago remains an open calibration question. In theory, the answer should be deflation.
The MetaHarness process
Tenex standardized its engineering skills into a structured set of documentation inside each repo, plus a CLI that agents use to drive the work. That covers planning, visualizations, client readouts, and the code itself. A canonical plan document sits at the top, though the enforced structure has loosened as models improved. Agents follow an opinionated convention for logging their decisions and activity, and everything is checked into Git.
The argument is that code alone is not sufficient memory. Teams need to know why a decision was made and what conversation led to it. Keeping that record durable in the repository, rather than in people’s heads, gives agents the context they need to act with more autonomy. The stated goal is for agents to drive more of the work, so humans no longer have to supply everything up front.
Models, tooling, and token economics
Opus 5.5 is currently the default driver for most of the team, covering what used to require two or three models: cheap and fast, very smart, and good at specific problems. Codex is still used for review and for computer-use-heavy work, with the two models working back and forth. Amp is the preferred interface, partly for its cross-model functionality and its cloud agents, which Alex describes as orbing out. Amp’s recent support for connecting a Claude Code subscription was described as cutting costs by roughly 80%. Alex argues that Anthropic’s cloud products aim at general collaboration, while Amp is opinionated and replicates local workflows.
Tenex’s ratio of token spend to employee spend is around 1:1 to 2:1. Alex compares this with the typical enterprise range of about 10 percent and with AI-native startups that reach 5 to 10x. Costs have come down since the period when the most expensive models were used for everything. The team now reaches for the largest model selectively. Early on there was worry about cost as headcount grew, but the team never embraced token leaderboards. The advice is that unconstrained budgets are when people run the experiments that teach lessons they can carry forward.
The discussion then turns to data-labeling companies such as Micro1, Mercor, and Handshake, where more than 90 percent of revenue comes from the labs. Founders expect enterprises to become the main buyers, which would require enterprises to fine-tune their own models. That only makes sense on quality or price. Quality is a hard bet because the frontier keeps moving. On price, per-task costs are falling quickly, which makes runaway token spend far less dangerous than it was six to eight months ago. A 10x smarter model rarely justifies its cost on routine work like moving cells in a spreadsheet.
The enterprise adoption gap
The speakers frame enterprise AI adoption as a sequence of milestones. Of roughly 100 enterprises Tenex is talking to, about 90 percent have rolled out Claude or Codex to employees. About 60 percent have invested in training and enablement, which Alex considers at least as important as the technology. Only about 50 percent have rolled out coding agents to their engineering organizations, far below the 70 to 80 percent that people guess. Agents running in production for non-engineering work sit at roughly 5 percent, especially among companies with more than a thousand employees. This is where the biggest diffusion gap lies.
Several forces explain the gap. FOMO drives adoption, and it is not inherently bad, but it pushes people toward whatever is most in the zeitgeist, which is why back-office agents are the most common request. Permissioning is another blocker. As every employee gets Claude or Codex, CIOs and CISOs struggle to control which data each agent can read, and they lack a gateway to act as a control panel. Tenex began by building bespoke gateways and now productizes its own.
A further obstacle is that you cannot automate a process that does not exist. When 40 people do roughly the same job in slightly different ways, a proactive background agent is hard to build. Tenex’s recommendation is to standardize the process first, then hand the work to an agent with defined steps. Without standardization, evals cannot be designed and the agent cannot know whether it is doing the job correctly.
Vertical AI and the future of SaaS
Alex and CJ say they are generally bearish on vertical AI applications such as Harvey for lawyers and OpenEvidence for doctors, even though those companies are currently winning. Their reasons are enterprise nuance and inertia, the expectation of one-of-one software, and the convergence of services and software as every software company builds forward-deployed engineering teams. They also question whether such businesses can keep software margins under that expectation, and why frontier labs would not simply absorb them.
The counterpoint is that proprietary data access is a real moat, and focused affordances matter. Legora and Harvey have customized UIs, such as tabular review, that serve their specific users. The Bloomberg terminal is cited as an example. The same market data is available elsewhere, but professionals pay for specialization and workflow integration. Superhuman is the analogy for a premium product on top of a commodity layer. The likely end state is that a small number of companies own the interface and other products plug into their agents through APIs and MCP servers, as Salesforce is pursuing.
Alex then describes his Kill My SaaS experiment, in which Tenex replaced a mid-tier SaaS product quoted at $40,000 a year for about $1,000 in cost. The tool is going into production for an upcoming conference. Their view is that rip-and-replace grows harder as you move upmarket. Salesforce and SAP are not at serious risk because the stakes of a mistake are so high. For mid-tier products, the question is whether the savings justify the cognitive load of maintaining non-core software. The offensive case is that customers who can iterate in minutes instead of months could build a far better product, though enterprises rarely behave that way.
AI, jobs, and agency
Alex argues that AI will augment and displace work rather than simply cut it. He points to reports that call centers are in decline and says this is devastating for people whose careers were built on that work. Tenex’s approach with clients is reallocation. The company first replaces a task, then looks at the next largest problem. Because organizations have effectively infinite demand, the goal is to move deep subject-matter experts toward cross-team and strategic work.
CJ pushes back on the framing that jobs are simply displaced. He argues it gives too little credit to individual agency. Reskilling is the job of both employers and employees, and people should ask how stable or replaceable their roles are over the next few years and act on the answer. He notes that many people do not have the luxury of enjoying their work, so the idea that AI will only remove unpleasant tasks is partly wishful.
New York’s AI ecosystem
Both speakers say the San Francisco AI bubble is larger and more self-referential than anything in New York. Tenex sees itself as a microcosm of that scene, since the people who self-select into the company and its partners are already AI-forward. The aim is to make New York a hub for AI diffusion, given how much commerce happens there and how large an impact diffusion could have on national GDP. AI billboards are starting to appear across the city, a sign that the culture is spreading. Alex’s model is arbitrage: bring San Francisco people to New York to meet customers and to get a reality check.
Hiring at Tenex
Tenex is hiring forward-deployed engineers, AI strategists who act as technical consultants leading client engagements, and especially engineers. The New York team is about 75 people. CJ’s current special projects include building the MetaHarness system to find the edges of what models can do and to give them better tools, as well as an internal platform for deploying background agents into enterprises to solve messy IT problems.