Agentic AI that heals systems that can't be replaced

Core operations at banks, insurers, hospitals, and government agencies still run on mainframes, COBOL, and FORTRAN rooted in the 1960s–70s.

These systems are as fragile as they are old, with tangled dependencies that mean they can't be stopped or rebuilt. So people have kept them running by physically learning the exact order to press buttons, how to ignore warnings, and rules like entering the same value twice.

Amazon's AGI Lab trains agents in reinforcement learning (RL) gyms and synthetic web environments modeled directly on this reality. The key is not just teaching successful flows, but having the agent learn the system's "true behavior," including failure patterns like delays, error states, re-renders, and partial saves.

For example, in a welfare application simulation, there was a case where an agent noticed on its own that it had added only one dependent, went back to correct it, and continued the flow. This capability is interpreted as a sign that the agent is beginning to understand the business rules and sequences hidden behind the interface, not just automating clicks.

From this perspective, the agent becomes a new interface layer placed on top of existing systems.

• Converting unstable UI into stable meaning: handles delays, retries, and partial failures as predictable patterns.
• Abstraction across systems: a single agent traverses workflows that were never designed to connect with each other.
• Incremental modernization: components can be changed bit by bit without overhauling the whole system.
• Preservation of organizational knowledge: the agent retains undocumented procedures and tacit knowledge.

Ultimately, agentic AI is less a technology that replaces old administrative, financial, and medical infrastructure, and more like a new API that runs on top of systems that can't be replaced. Innovation doesn't only come from discarding what exists—it can also start from keeping fragile systems alive as-is and turning them into a stable platform.