The atmosphere at the Code with Claude event in London was not one of cautious assistance, but of aggressive autonomy. While the tech world was focused on Google I/O on May 19, Anthropic was quietly signaling a paradigm shift in how software is built. Ravi Trivedi, an engineer at Anthropic, summarized the company's new philosophy with a simple, provocative phrase: Let it cook. This is not about a better autocomplete or a more helpful chat interface. It is about getting the human developer out of the way entirely, moving from a world where AI suggests code to one where AI manages the entire lifecycle of a feature, from conceptualization to deployment.
The Architecture of Autonomous Engineering
The transition toward this autonomous future was paved by the rapid release cycle of Anthropic's latest models. Following the launch of Claude 4.6 in February and Claude 4.7 in April, the company has moved beyond the experimental phase of AI coding. While the original Claude 4 served as a limited assistant, the 4.7 iteration is designed for full delegation. This shift is already manifesting in the workflows of global enterprises. Spotify and Delivery Hero have integrated Claude Code into their internal processes, fundamentally restructuring how their engineering teams operate. Meanwhile, startups like Lovable, Base44, and Monday.com are pioneering a movement known as vibe-coding. In this new approach, the rigid adherence to detailed technical specifications is replaced by an intuitive, iterative flow between the developer and the AI, where the application is shaped by interaction rather than a static blueprint.
The most striking evidence of this autonomy exists within Anthropic itself. Jeremy Hadfield, an engineer at the company, revealed that a vast majority of Anthropic's internal software was written by Claude. Most tellingly, the code for Claude Code itself was authored by the model. This creates a recursive loop of self-improvement that challenges the traditional role of the software engineer. During the London event, a survey of attendees revealed a startling trend: many developers admitted to approving and deploying pull requests without reading a single line of the generated code. The traditional safety net of human peer review is not just being streamlined; it is being bypassed entirely in favor of speed and trust in the model's output.
The Dreaming Engine and the Productivity Paradox
The primary obstacle to autonomous coding has always been the loss of context. Historically, AI agents suffered from a form of digital amnesia, requiring developers to re-explain constraints or past errors every time a new session began. Anthropic addresses this through a system called Dreaming. This is not a metaphorical term but a specific memory management architecture. When an agent performs a task, it records useful insights, edge cases, and solutions into a set of Notes. These notes are not merely stored; they are shared. When a subsequent agent is assigned to the same codebase, it immediately references these notes to avoid repeating previous mistakes and to align with the existing architectural logic.
This Dreaming system evolves from simple storage into integrated analysis. The system periodically reviews these accumulated notes to identify recurring patterns and systemic issues across different tasks. This transforms a collection of individual agents into a collective intelligence that grows more specialized to a specific codebase over time. Supporting this is a mechanism called self-prompting. Rather than waiting for a human to provide a command, Claude Code now generates its own internal prompts. It defines its own sub-tasks, executes them, and runs its own tests. If an error occurs, the AI enters a self-correction loop, fixing the bug and verifying the solution before the human developer is even notified that a problem existed.
However, this leap in capability introduces a dangerous productivity paradox. While Claude 4.7 operates at the level of a mid-level engineer, the human capacity to review code remains static. We are seeing a massive divergence between the speed of AI production and the speed of human verification. On platforms like Reddit and Hacker News, developers are reporting a new kind of burnout: the burden of reviewing an overwhelming volume of AI-generated code. This creates a systemic risk where managers, chasing productivity metrics, encourage the deployment of unverified code. The result is a hidden accumulation of technical debt and potential security vulnerabilities. An AI can produce syntactically perfect, elegant-looking code that nonetheless contains deep architectural flaws or security holes that only a senior human eye would catch. By removing the human from the loop, the industry is trading long-term stability for short-term velocity.
This tension is palpable even within Anthropic. Technical managers have expressed exhaustion over the sheer volume of code their teams are now producing. The development cycle has shrunk, but the cognitive load of oversight has exploded. As the tool evolves to require less human intervention, the gap between the high-level architect and the implementation layer widens. The risk is that the industry may lose the very expertise required to troubleshoot the systems that the AI is building, creating a future where we possess complex software that no living human fully understands.
The era of the developer as a writer of code is ending, replaced by the developer as an orchestrator of intelligence.
