Claude Code’s Creator Says ‘Taste’ Isn’t Humanity’s Last Moat – What Do Companies Hire When Engineers Stop Coding?

Claude Code’s creator challenges the idea that human "taste" is the final defensive moat. Boris Cherny, a technical member at Anthropic, observes that models are rapidly learning what to do, raising the question of what remains uniquely human.

How Claude Code was born

In late 2024, Cherny joined Anthropic’s Labs team, tasked with exploring future product forms rather than maintaining existing ones. The team felt models already surpassed current AI programming tools, which were limited to autocomplete or Q&A assistants, and saw a gap for a true "Coding Agent." They decided to make the model the primary developer, building a product around it. Early versions of Claude Code could only handle 10‑20% of tasks, requiring extensive manual coding.

Why Anthropic focuses on coding

Anthropic’s core mission is AI safety, not commercial gain. Coding provides a clear, binary feedback loop—code either runs or fails—making it an ideal real‑world testbed for model behavior. The abundance of open‑source code also offers rich training data, allowing precise evaluation of model capabilities.

What made Claude Code suddenly stronger

According to Cherny, the decisive factor was the underlying model improvements. Upgrades from Sonnet 4 to Opus 4 and then Opus 4.5 directly boosted Claude Code’s performance. While the product added new interfaces (CLI, desktop, mobile, Slack, GitHub) and features like Plan Mode, Cherny views these as incremental; the real leap came from the model itself.

Productivity impact at Anthropic

Internal metrics show that after wide adoption of Claude Code, each engineer’s code output roughly tripled, with actual growth now exceeding that figure. New hires now ramp up in two days instead of weeks because they can query Claude for knowledge that previously required senior engineers. Knowledge transfer costs have collapsed, as hidden expertise is encapsulated in the agent.

From punch cards to Vibe Coding – raising the abstraction level

Cherny likens the evolution of software development to a series of abstraction jumps—from punch cards to assembly, Fortran, Java, Python, and now AI agents. He argues that today’s shift mirrors past jumps: humans are simply moving to a higher level of abstraction where the model handles low‑level coding.

Engineers no longer write code – what do you hire for?

When asked how Anthropic evaluates engineering candidates in a world where code is largely generated by models, Cherny says the company now seeks "Generalists"—people who can span design, data analysis, product thinking, and engineering. Traditional narrow roles are dissolving; even designers and finance staff write code. The title "Member of Technical Staff" (MTS) reflects this fluidity, emphasizing equal footing and focusing on the ability to turn ideas into reality rather than seniority.

Advice for founders

Cherny recommends giving teams abundant tokens for experimentation and deliberately staffing projects with fewer people, trusting the model‑augmented workflow to automate much of the work. This token‑first approach raises upfront costs but dramatically lowers ongoing expenses, akin to pre‑compiling code.

Is "taste" being eroded?

He observes that personal coding preferences (e.g., functional programming) lose relevance as the model produces effective code regardless of style. The notion that "product taste" is the final alpha is also fading; models now generate the majority of ideas, with only a minority initially promising.

Core values remain

When asked what humans still have that models lack, Cherny answers "values"—the ability to do the right thing, not just do things right. He likens teaching models to teaching children: instilling good existence and ethical judgment.