Frontier Labs Reorganize Around Recursive Self-Improvement as Agents Drive 30,000 Daily Code Merges

• Frontier labs have restructured into "AGI organizations" where AI agents execute the majority of research and engineering tasks, creating a recursive self-improvement loop that other industries cannot replicate.
• At Anthropic, over 80% of merged code lines were written by Claude as of May, and daily merged changes increased from 2,500 in May to more than 30,000 in September.
• Internal AI usage costs have surged, with Anthropic's internal Claude use growing 250x from January to September and passing monthly personnel costs (excluding stock compensation) in April; OpenAI researchers' daily inference usage reached a median of over $600.
• Organizational roles are merging as engineers, product managers, and designers collaborate on agent-driven workflows, while infrastructure shifts to support agent permissions and high-volume code shipping.

Operational Metrics

• OpenAI's research organization recorded 3.1 agent workdays for every human workday as of mid-August.
• Anthropic's Claude Code users approved 93% of permission prompts, leading to a shift from monitoring actions to limiting capabilities via sandboxes and risk classifiers.
• A specific automated research experiment at Anthropic yielded a 0.5-point gain on production infrastructure, a result the author notes is within noise and highlights the current fragility of AI-proposed hypotheses.

Economic and Structural Shifts

• Compute allocation is the primary bottleneck, with research compute viewed as a bet on future models while customer compute drives current revenue; over 80% of Anthropic's Q3 revenue came from models released in the previous six months.
• Hiring criteria are shifting to evaluate "AI-driven impact," with companies like Google DeepMind allowing coding agents in engineering interviews.
• Non-frontier companies, such as the game studio Spawn, are experimenting with agent-heavy workflows, spending roughly 30x their payroll on tokens to ship code hourly, though they lack the internal model loop of frontier labs.