Former Anthropic Engineer Claims Over 90% of Engineers Use Self-Improvement Loops
A former Anthropic engineer reportedly shared that more than 90% of the company's engineers utilize self-improvement loops in their software development workflows. The claim highlights how these iterative, agentic loops are actively used within Anthropic to build and refine AI systems. This reveals the widespread adoption of agentic workflows and recursive self-improvement paradigms in top-tier AI labs, signaling a shift from manual coding to AI-assisted iterative development. It suggests that self-improving code loops are becoming a standard practice for accelerating AI engineering. While the source tweet lacks deep technical specifications, it points to a shared methodology for constructing these loops in practice. In agentic AI, self-improvement loops typically involve models generating code, testing it, and refining it based on feedback without constant human intervention.
## BACKGROUND
Self-improvement loops in agentic AI refer to systems where AI agents recursively evaluate and upgrade their own code, prompts, or architectures to achieve better performance. Agentic workflows extend traditional automation by enabling AI models to act as autonomous agents that plan, use tools, and execute multi-step tasks.