The Reading Bottleneck: When AI Writes More Code Than Humans Can Review

For decades, the speed of software development was bottlenecked by how fast a human could type. Today, it is increasingly bottlenecked by how fast a human can read.

At a recent "Code with Claude" developer event hosted by Anthropic in London, a striking reality emerged about the modern tech industry. When asked if they had recently shipped a software update written entirely by AI without actually reading the code themselves, nearly half of the developers in the packed room kept their hands raised.

This willingness to blindly trust machine output highlights a radical shift in how software is built. Anthropic is actively encouraging this hands-off approach. Their new operational philosophy for developers is simple: "Let it cook." Instead of a human prompting an AI, finding an error, and prompting it again, Anthropic wants its Claude AI to prompt itself. The goal is for the AI to encounter a bug, test a hypothesis, and tweak the code autonomously until it works, entirely out of the human developer's sight.

To facilitate this, Anthropic recently rolled out a feature called "Dreaming." Claude Code agents can now write internal notes detailing the quirks and specific requirements of a given task. The Dreaming system later consolidates these notes, allowing the AI to spot patterns and effectively "learn" a company's unique codebase over time, much like a newly hired engineer getting up to speed.

Unsurprisingly, major tech players like Spotify and Delivery Hero are eagerly reorganizing their workflows around these capabilities. Yet, beneath the corporate enthusiasm, a grassroots backlash is brewing.

On developer forums like Hacker News, engineers are sounding the alarm about the hidden costs of AI productivity. As managers push for faster output using AI, developers find themselves drowning in a flood of generated code that they must somehow review and verify. The mental toll of deciphering vast amounts of machine-written logic is leading to widespread exhaustion. Even Katelyn Lesse, an engineering lead at Anthropic, admitted that some of their own technical managers are exhausted just trying to keep up with the sheer volume of code their teams now produce.

Furthermore, critics worry that relying too heavily on AI will erode fundamental human coding skills and introduce subtle security vulnerabilities that overworked reviewers might miss.

Anthropic makes no secret of its endgame: building an AI that can essentially build itself. But as we transition into an era where machines write the vast majority of our software, the tech industry must confront a difficult paradox. We have successfully automated the creation of code, but we have yet to figure out how humans can safely and sustainably supervise a tireless digital workforce.

Key Points

• Many developers now regularly deploy AI-written code without reviewing it manually.
• Anthropic's 'Let it cook' philosophy encourages AI to autonomously test and debug its own work without human intervention.
• A new 'Dreaming' feature enables Claude to leave notes for itself, helping the AI learn and master specific codebases over time.
• The massive surge in AI-generated code is causing severe review fatigue among developers and engineering managers.
• Experts warn that over-reliance on autonomous coding tools could degrade human skills and introduce security risks.

Why It Matters

As AI transitions from a coding assistant to an autonomous developer, it forces the tech industry to rethink how software is reviewed, secured, and maintained when machines outpace human oversight.

Sources:

• Anthropic’s Code with Claude showed off coding’s future—whether you like it or not — MIT Technology Review - AI