Summary

10x’s founding insight is that AI-native engineering makes hourly compensation perverse: faster builders can earn less as they create more value. After a 90% engineering downsizing forced Arman Hezarkhani to rebuild Parthean’s product-and-engineering process around AI, production-ready output “10x’d.” The company now pays for story-point output, seeking “unlimited upside” for exceptional engineers.
The model could produce seven-figure individual compensation in cash from story-point compensation alone. Arman Hezarkhani says 10x will “probably” have more than one engineer earn $1 million next year, and is “very likely” to have more than a handful cross that threshold.
Story points remain gameable, so 10x treats incentives and hiring—not measurement—as the real control system. It selects for people who are “selfish but long-term selfish,” then pairs engineers with technical strategists rewarded for NRR, retention, and account growth. That internal tension creates a final quality gate before clients see the work.
The best evidence is prototype compression from quarters to weeks—or hours. 10x put several quantized retail-vision models onto Raspberry Pi 4s, Jetsons, and Nanos in two weeks; built a mobile app that reached No. 20 globally in a month; and turned a rejected sales pitch into a working fitness, health, and nutrition coach in roughly four hours. Hezarkhani still cautions: “We’re not claiming that we’re magical beings.”
10x is currently human-bound, not agent-bound, making elite technical recruitment the binding constraint on growth. Its deliberately “unreasonably difficult” take-home draws no response from about 50% of candidates, but lets successful applicants finish the process in as little as one week.
The stack is optimized for agent feedback loops, while tool selection remains deliberately fluid. Shared TypeScript schemas give agents constraints and useful errors; Claude Code, Codex, and Cursor are chosen according to the task and even the day. Swyx challenges the anecdotal comparisons and asks about comprehensive evals, then frames the counterpoint as a samurai’s sword: once tools are good enough, feel and fit matter; Hezarkhani agrees.
Fully autonomous engineering may hinge less on raw intelligence than on preventing small errors from compounding. Hezarkhani names model intelligence as the main blocker; Alex Lieberman points to context engineering; 10x engineer Dan reframes the issue as “controlling entropy,” because even a 1% error rate can accumulate until an autonomous loop derails. On MCP, Hezarkhani calls it “a three-letter word for API,” while Lieberman says it is useful but criticizes hype and outsized fundraising around the label and defends the broader protocol as more than API wrappers.

Deep dive

1. A forced 90% downsizing exposed AI’s real engineering leverage

Lieberman met Hezarkhani in 2020 after investing in Parthean, an AI financial-tools business that moved from consumers toward financial advisors and RIAs. The pivotal conversation came roughly nine months before this interview: Hezarkhani had reduced his engineering organization by 90% and could no longer preserve its old product-and-engineering process.
Necessity forced an AI-first rearchitecture. Hezarkhani reported that production-ready software output then “10x’d”—a claim Lieberman initially resisted because, despite life-changing experiences with ChatGPT and Grok, he had never personally seen that degree of leverage.
The resulting economic thesis: an engineer producing ten times the throughput cannot credibly quote “a thousand bucks an hour,” while hourly billing perversely rewards slower work. 10x therefore pays for output and asks how exceptional engineers can receive “unlimited upside.”

2. Story points work only when incentives extend beyond the sprint

Swyx’s core objection was direct: what is a unit of software output—a PR, a story point—and won’t “what gets measured” simply get gamed? Hezarkhani conceded the system is exploitable; equating every line of code with more points would mechanically inflate compensation.
His defense rests on repeated relationships. An engineer can manipulate today’s points, but poor work makes clients churn and ends the opportunity; 10x therefore hires people who are “selfish but long-term selfish,” alongside builders who simply enjoy writing code with strong peers.
Lieberman added an organizational check: each engagement has an AI engineer and a technical strategist. The strategist is rewarded for NRR, retention, and account growth, then provides the final approval on the engineering plan before a sprint begins—“two people at odds with each other in a healthy way”—creating the last quality defense before clients see anything.
10x had not yet faced a client dispute over point assignment or alleged sandbagging, though Lieberman stressed that the company is young. Swyx’s interpretation: point allocation becomes political when delivery goes badly; when it goes well, everyone keeps “steaming ahead.”

3. Prototype speed is becoming both delivery advantage and sales weapon

The most technical example involved a retailer-technology client using a Raspberry Pi 4 in stores. 10x combined off-the-shelf and internally trained models, quantized them, and ran several in parallel on Raspberry Pi 4s, Jetsons, and Nanos to produce heat maps, identify queues, assess shelf-stocking needs, and support theft detection through body analysis.
The prototype took two weeks, versus what Hezarkhani said previously would have required several quarters from a robust engineering team. He kept the limitation intact: this remained a research project requiring prolonged accuracy and metrics work, not proof of “magical beings.”
Fast prototyping also changes sales. After a fitness influencer rejected 10x as too early—and 10x lacked a built-in design team—an engineer built a working personalized fitness, health, and nutrition coach in roughly four hours; the app has not launched, but 10x moved to first place on the prospect’s list to do the build.

4. Structured code helps agents, but no coding agent stays champion

10x’s default is TypeScript across front and back end, with shared types and schemas. The attraction is JavaScript’s flexibility plus TypeScript’s constraints and error messages, which let Claude Code, Cursor, or another agent run, inspect failures, and continue iterating.
Hezarkhani rejected the idea of a favorite agent “of the year or of the month or even of the week.” The team might prefer Claude Code at 4:42 today, then find Codex superior for particular activities tomorrow.
Swyx pushed back that this is anecdotal and vulnerable to “the luck of the draw” without comprehensive evals. He supplied the experiential counterpoint: once coding agents are sufficiently capable, a practitioner’s fit matters—how an agent collaborates and whether it writes code in the engineer’s preferred style. Hezarkhani agreed.
Despite the tooling leverage, Hezarkhani called 10x “human-bound 100%.” The immediate bottlenecks are finding enough excellent engineers and building processes that preserve delivery quality; building 10x’s own technology is a longer-term ambition.

5. Autonomy fails when errors compound into entropy

10x retains take-home interviews after many peers abandoned them, but makes the assignment “unreasonably difficult”; about 50% do not respond. The payoff is a short process—two calls, the take-home, review, and one or two final meetings, potentially completed within a week.
Asked what blocks a fully autonomous senior engineer, Hezarkhani proposed model intelligence as the main blocker, noting that existing training has generalized more readily to Python and Django than to full back-end distributed services. Lieberman instead emphasized context engineering—getting the right information into the LLM and directing attention correctly. Engineer Dan sharpened the problem to “controlling entropy”: at 99% accuracy, the residual 1% can multiply through an autonomous loop until accumulated error derails the agent.
The MCP exchange exposed a difference in emphasis. Hezarkhani called MCP “a three-letter word for API” and took a nonjudgmental sociological view of how technical communities invent terminology. Lieberman said MCPs are useful but criticized hype and outsized fundraising around a renamed concept, while defending the broader protocol as more than API wrappers. He also argued that real debate reveals more than an unchallenged talk.

No Priors Ep. 141 | With Sunday Robotics Co-Founders Tony Zhao and Cheng Chi Ari Emanuel’s "Anti-AI" Bet on Live Entertainment