AI in NYC: How Tenex Builds an AI-Native Engineering Team — Alex Lieberman & CJ Hess

Summary

• The most tradeable data point is Alex Lieberman’s enterprise adoption funnel: of about 100 enterprises 10X talks to, 90% have rolled out a general-purpose LLM (Claude or Codex), about 60% provide training and enablement, and about 50% have deployed coding agents — versus the 70–80% many AI Engineer attendees guess. Agents in production for non-engineering work fall to about 5%. Alex describes this as the biggest gap between raw intelligence and enterprise usage, underscoring 10X’s thesis that the final mile of diffusion is hard.
• 10X’s output-based compensation thesis has largely held up, but the model has needed nuance: Alex says the feared gaming of story points and shipping of bad code have not been concerns, with AI strategists instead incentivized for customer retention and growth. The biggest compensation challenge is accounting for engineers’ non-coding work, such as architecture and client meetings. The earnings-capacity gap between the highest- and lowest-paid engineer is about 4X, not 10X. They recalibrate story-point criteria at least quarterly; one speaker says that, in theory, points should deflate as models improve.
• CJ Hess estimates engineer output has roughly doubled over the period from the “Opus 4.5 moment” in December to Opus 5.5 today, which he calls a “magnificent model.” Opus is fast, cheap, and covers the range for which he previously used two or three models. The team often pairs it with Codex for review; 10X’s “meta-harness” keeps plans, decisions, and agent logs in the repo, so context is “never in someone’s head.”
• Token-cost anxiety has eased, but was a concern as headcount grew. CJ estimates 10X’s raw token spend is between 1:1 and 2:1 relative to employee spend. swyx’s benchmarks include $10–$200 per employee at many companies and token spend ranging from roughly equal to employee spend up to 5–10× at AI startups. CJ says falling per-task costs have made it harder to “token max” and seriously hurt the company than six to eight months ago. Alex’s spreadsheet test: “Would a 10-times-smarter model really justify the cost of moving some cells in a spreadsheet? Probably not.”
• The episode’s sharpest disagreement: Alex is generally bearish on vertical AI such as Harvey and OpenEdmunds, arguing that enterprise buyers expect one-of-one solutions, which pushes vendors toward FDE teams and puts software margins under pressure. swyx pushes back: specialized data, affordances, and integrations — including Bloomberg’s value beyond access to market data — can sustain vertical products. Alex points to Bloomberg’s social network; swyx says the broader software market will grow, leaving companies to compete for share.
• Alex argues that Salesforce and SAP are hard to rip out because a failed replacement could put hundreds of millions or billions of dollars at risk; mid-tier SaaS is more case by case. swyx’s “Kill My SaaS” example: a vendor quoted $40,000 a year, while the replacement cost $1,000 and is headed to production. The case for rebuilding is not just cost: swyx says the experience could be much better when the customer can iterate in minutes rather than months.
• Two recurring enterprise problems are deploying back-office agents and managing data permissions. Alex says permissioning and RBAC have become an “absolute nightmare for CIOs and CISOs,” and 10X has productized its gateway solution. CJ’s prerequisite framing: “You can’t automate a process that does not exist.” When 40 people do roughly the same function differently, standardize the process before handing work to an agent or building evals.
• On job displacement, Alex emphasizes people’s agency; CJ makes the harder point that calling AI a way to replace only work people dislike is “a little bit of cope.” CJ cites a16z on call centers declining, while 10X frames deployments as reallocating people to the next problem. On geography, CJ sees a smaller AI bubble in New York than in San Francisco, while Alex argues New York may be as large or larger commercially and can help diffuse AI.

Deep dive

1. Paying engineers for output, not time, has held up — with a 4X dispersion, not 10X

• Alex reintroduces 10X as “an applied AI partner for some of the biggest companies in the world,” helping them go from AI apps to AI-native. The premise is that AI is “raw intelligence” and the hard problem is the final mile of diffusion. Eighteen months in, the business has 75 people, 90% of them engineers or AI strategists; customers are split between Fortune 2000 companies and mid-market companies doing $100–$500 million in revenue.
• CJ’s origin story is the compensation thesis in miniature: as a consultant billing hourly, “I just felt my output getting better and faster, and ultimately getting paid less.” He took the bet, back in the Sonnet 3.5 / Gemini 2.5 Pro era, that output-based pay would let him earn more as AI increased his output.
• The classic objections — sandbagged story points and barely-checked-box code — “haven’t been concerns at all,” Alex says, because AI strategists are incentivized for customer retention and growth, in a way he describes as “perfectly at odds with engineers.” The actual hard problem has been compensating engineers for time spent architecting or meeting clients. Alex puts the earning-capacity difference between the highest- and lowest-paid engineer at about 4X — prompting swyx’s “Oh, not 10X?”
• Story-point criteria are recalibrated at least quarterly. One speaker says that, in theory, points should deflate as models improve; CJ describes the balance between rewarding output and recalibrating the value created. CJ’s forward view: more value lies in “sitting down with a customer, hearing everything that they can brain dump, and then just providing that to the model.”

2. The meta-harness: context in the repo, output roughly doubled over the period

• CJ describes 10X’s engineering system as structured documentation plus a CLI the models use to drive work: canonical plan documents, agent decision logs, and visual readouts. The material is checked into Git. On the docs-versus-code debate, he says you do need to know why a decision was made; the durable context is “never in someone’s head. It’s always in the repo.”
• One measure of whether the processes are working is average story points per engineer per month. CJ estimates output has roughly doubled over the period from the December “Opus 4.5 moment” to today, when the team is using Opus 5.5. He says models began taking on long-running tasks around that point; the difference in what engineers are building and shipping is “insane.”

3. Opus 5.5 has reduced the model rotation, and token-cost concerns have eased

• The office “vibe check” is Opus 5.5. CJ says he doesn’t particularly use Fable and thinks many people are defaulting away from it. Opus is “fast, it’s cheap,” and covers the capabilities for which he previously used two or three models. The team still reaches for Codex for computer-use work and often uses it to review Opus’s work. CJ says he is still AMP-maxing, using AMP’s Orbs cloud setup and its ability to connect a Claude Code subscription; swyx says that can cut costs by about 80%. CJ prefers AMP’s cross-model functionality to what he could get with Claude Code, Claude, and Claude Tag.
• On swyx’s token-spend-to-employee-spend benchmarks, CJ estimates 10X’s raw token spend at between 1:1 and 2:1. swyx says many companies spend $10–$200 per employee, while AI-startup token spend can range from about equal to employee spend up to 5–10×. CJ says costs were a concern as headcount grew, but are now lower: Opus 5.5 and subscription access have helped, and per-task costs have fallen. He says it is harder to “token max” and seriously hurt the company than it was six to eight months ago.
• CJ raises a question about data companies MicroOne, Mercor, and Handshake: he says more than 90% of their revenue is from labs, while some founders expect enterprise demand to grow in the next few years. That expectation depends on enterprises fine-tuning their own models, which CJ considers a hard bet if the frontier keeps advancing and token costs keep falling. Alex’s caveat is that if an Astra 6.5 or Fable 5.5 delivers the best results at a high price, spending could trend back up. He also cites Muse Wonder as frontier and “dirt cheap.”

4. The adoption funnel: 90% have an LLM, 5% have a non-engineering agent in production

• Alex’s enterprise milestone race: of about 100 enterprises 10X talks to, he estimates 90% have rolled out Claude or Codex; training and enablement sits at about 60%; coding agents at about 50%, versus the 70–80% many AI Engineer attendees guess; and agents in production for non-engineering work at about 5%. Alex calls that last drop the biggest diffusion gap between raw intelligence and enterprise usage.
• Alex says the wave is driven by a small group who understand what the technology can do and a much larger group driven by FOMO. That momentum can also lead enterprises to latch onto the agent zeitgeist. One common request is a back-office agent; another major issue is permissioning and RBAC. With employees using Claude or Codex and needing access to company data, Alex says, it has become “an absolute nightmare for CIOs and CISOs.” 10X first built bespoke solutions and has since productized its own gateway.
• CJ’s prerequisite warning: “You can’t automate a process that does not exist.” When 40 people do roughly the same function in different ways, it is harder to build an agent that handles the work in the background. Standardize the process first, CJ says, so the agent can be evaluated and the work handed off; the goal is for people to become more like agent managers.

5. Bearish on vertical AI, while swyx argues specialized products can endure

• Alex says he is generally bearish on vertical-specific AI such as Harvey for lawyers and OpenEdmunds for doctors. From 10X’s vantage point inside enterprises, he sees nuance, inertia, and many systems and people to coordinate. Buyers increasingly expect one-of-one solutions, while software companies build FDE teams; Alex questions how vertical software can meet those expectations at software margins, and why frontier models would not “gobble them up.”
• swyx pushes back: “So far you’ve been wrong.” He points to data ChatGPT may not have, as well as specialized affordances and interfaces — citing Harvey’s tablet review and Legora. His Bloomberg example: although he could access the same information from Nasdaq and the New York Stock Exchange, Bloomberg brings consolidated, specialized information, peer usage, and integrations. Alex counters that people also pay for Bloomberg’s social network and DMs.
• CJ distinguishes a new AI-built app from a vertical product that is mainly an agent doing work. He thinks vertical agents without a strategic compliance or data moat will struggle as models improve, while apps and tools that change the process and put an agent on top can still be successful. swyx adds that users may prefer to use an agent they already have and plug in their data; he notes Salesforce is embracing that path with APIs and MCPs.
• swyx’s “Kill My SaaS” example sharpens the mid-tier question: the vendor quoted $40,000 a year, the replacement cost $1,000, and it is going into production for the next conference. Alex says Salesforce and SAP are hard to rip out because a failed replacement could put hundreds of millions or billions of dollars at risk. Mid-tier SaaS is case by case: the question is whether the opportunity justifies allocating engineering resources and taking on the cognitive load of maintaining non-core software. 10X has built its own CRM and OS. swyx’s point is that rebuilding can be about offense, not just cost: the experience may be much better when the customer can iterate “in minutes and not months.”

6. Displacement without easy answers, and New York as the commercial pond

• CJ says AI will augment and change work, and will also displace some of it. Citing an a16z post, he says call centers are declining as AI handles more of the work. 10X frames agent deployments as reallocation: once one task is solved, there is another problem to tackle, and subject-matter experts may be able to work across teams or focus on more strategic questions.
• Alex argues that job displacement is often discussed as something that happens to people, without giving enough credit to their agency. CJ adds that describing AI as replacing all the work people dislike is “a little bit of cope”: many people cannot choose work based on whether they enjoy it. He says both employers and employees have a role in reskilling, and people should assess the stability or replaceability of their jobs and take action where possible.
• CJ describes New York’s AI bubble as smaller than San Francisco’s, though 10X’s team and partners give him a view of a highly engaged local AI community. He says 10X wants to grow that community and make New York a hub for AI diffusion, given the amount of commerce there and the potential impact on GDP. Alex characterizes New York as a smaller pond for AI but as large or larger for commercial value. swyx says he holds AI Engineer in New York to bring San Francisco people to meet customers and get a reality check; Alex calls that an interesting arbitrage.
• CJ says his special-projects work includes building 10X’s internal meta-harness and platform, finding ways to give models tools that improve their capabilities, and helping deploy background agents in enterprises.