Anthropic's Multi-Agent Graphs Become Standard as Self-Improving Loops Hit 80% Engineering Adoption

Anthropic engineers have crossed a tipping point: 80% of the team now relies on self-improving loops, with graph-based agent orchestration becoming the new standard for building complex AI systems. In a workshop this month, Anthropic revealed that graph engineering—orchestrating multiple specialized agents through explicit task dependencies rather than single-agent loops—is reshaping how they build and iterate on AI infrastructure. The shift reflects a maturation in agent architecture beyond simple iterative refinement to what researchers call "experience graphs," persistent data structures that record exploration attempts, outcomes, and reusable knowledge across sessions.