How agentic AI helps heal the systems we can’t replace

Amazon's AGI Lab is pioneering a new approach to modernizing the world's most critical, yet brittle, software infrastructure. Instead of attempting to replace decades-old systems written in languages like COBOL and FORTRAN—the backbone of finance, travel, and government services—they are training AI agents to master them. These agents learn by navigating high-fidelity simulations that capture the real-world idiosyncrasies of legacy systems: mandatory windows that encode sequencing rules, fields that must be entered twice, and warnings that must be ignored. The goal is not to rebuild but to understand and interface with the accumulated layers of technology that are too vital to take offline.

This agentic AI effectively becomes a 'universal API,' sitting between users and the complex, archaic backend. It learns the institutional 'folklore'—the undocumented knowledge human operators pass down about which buttons to click and in what order. By managing these eccentricities, the agent provides a clean, modern interface to perform tasks while preserving the stability of the underlying, unchangeable systems. This represents a fundamental shift in enterprise modernization, focusing on healing and navigating existing infrastructure rather than attempting risky, multi-billion dollar replacement projects that often fail.