Beyond Job Cuts: Anthropic Warns That AI Recursion Is the New Frontier

As Anthropic shifts focus from labor market displacement to the risks of self-improving systems, internal data reveals that Claude now writes the majority of the company's own software.

For much of 2026, the public narrative surrounding Anthropic centered on a grim economic forecast: massive displacement of white-collar professionals. CEO Dario Amodei spent months warning global leaders—from Davos to the India AI Impact Summit—that unemployment could swell to 20% as finance, law, and coding sectors faced automation. Yet, a sprawling 10,000-word paper released by the company now suggests that job losses, while significant, may be a secondary concern. The new focus is "recursive self-improvement," a scenario where systems become capable of designing their own successors with minimal human oversight.

The Shift to Recursive Systems

The paper, When AI builds itself, authored by the Anthropic Institute’s Marina Favaro and Jack Clark, argues that the most pressing danger is the loss of human control over the development lifecycle. While the industry previously viewed the automation of entry-level coding tasks as the primary milestone, Anthropic’s internal data indicates that the goalposts have moved. The company notes that if the gap toward autonomous, self-training systems closes as rapidly as current metrics suggest, the focus must shift immediately toward the rigorous monitoring and safety architecture of these systems.

Claude as the Architect

Evidence of this transition is embedded deep within Anthropic’s own operations. As of May 2026, over 80% of the code merged into the company’s codebase was authored by Claude. This is a dramatic escalation from early 2025, when that share sat in the low single digits. Because the company’s engineers now ship eight times as much code per quarter as they did in the 2021–2025 period, the sheer velocity of development has fundamentally changed the nature of their work. Researchers polled in March 2026 reported they were producing roughly four times as much output using the internal Mythos Preview model than they would have without such assistance.

The potential for these systems to execute complex, long-term tasks is accelerating. In one instance, Claude shipped over 800 fixes in April 2026, resolving a class of API errors that would have taken a human engineer an estimated four years to complete. Furthermore, external evaluators at METR found that the Mythos Preview model could operate independently for at least 16 hours. With the duration of reliable task completion doubling every four months, the boundary between a tool that assists a coder and a system that engineers its own evolution is blurring.

Economic Implications

This pivot presents a complex dilemma for investors and policymakers. While the initial valuation of the company—pegged at $900 billion—was predicated on the massive potential to capture the global wage bill, the new warnings suggest that the technological risks may outpace regulatory frameworks. By documenting how Claude now contributes to the very infrastructure that sustains it, Anthropic is signaling that the era of human-led software design may be nearing an inflection point, forcing a reconsideration of what it means to "work" in a field that is increasingly building itself.