Anthropic Engineers Scale From Self-Improving Loops to Agentic Graphs

Anthropic is witnessing a fundamental shift in how its engineers build AI systems, with adoption of self-improving agent loops reaching critical mass and now transitioning wholesale toward graph-based architectures for coordinating multiple agents. In a 2026 engineering update, an Anthropic engineer revealed that 90% of the company's engineers were already running self-improving loops, and the entire organization is now moving toward agentic graphs—a more sophisticated approach to orchestrating multiple AI agents as interconnected systems rather than isolated loops.