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

2026-03-16 · Source: Amazon Science homepage · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Intermediate, long

Summary

Amazon's Artificial General Intelligence (AGI) Lab is developing agentic AI to modernize and manage critical legacy systems that are often slow, brittle, and outdated. These systems, built decades ago with technologies like COBOL and FORTRAN, underpin essential services such as finance, insurance, and public administration. Instead of attempting costly and often failing replacements, Amazon's approach involves training AI agents on high-fidelity simulations of these systems, including their quirks, delays, and error states. By learning the "real semantics" of these complex, layered architectures, agents can act as a universal API, providing a stable interface over unstable UIs and preserving institutional knowledge that is rapidly disappearing as original developers retire. This method enables incremental modernization and cross-system abstraction without requiring systems to be taken offline.

Key takeaway

For CTOs and VPs of Engineering grappling with aging, mission-critical legacy systems, this agentic AI approach offers a viable path to modernization without the prohibitive risks and costs of full replacement. Your teams can leverage AI agents to create a stable, programmatic interface over existing infrastructure, preserving operational knowledge and enabling incremental updates. Consider piloting agent-based solutions for workflows where human institutional knowledge is critical and traditional API integration is impractical.

Key insights

Agentic AI can modernize critical legacy systems by learning their real-world behaviors and acting as a stable, synthetic API.

Principles

• Innovation can grow from existing systems, not just replacement.
• System logic is revealed through friction and failure modes.
• Agents can preserve undocumented institutional knowledge.

Method

Amazon trains AI agents in reinforcement learning (RL) gyms, which are synthetic environments designed to reproduce the full spectrum of legacy system behaviors, including flaws, inconsistencies, and delays, to teach agents how to navigate and recover from failures.

In practice

• Use RL gyms to simulate complex, real-world system behaviors.
• Train agents on system failures to understand deeper logic.
• Implement agents as a stable interface over legacy UIs.

Topics

• Agentic AI
• Legacy System Modernization
• Reinforcement Learning
• Synthetic APIs
• Enterprise AI

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Building Self-Healing Software Systems with Multi-Agent AI Architectures — Multi-agent AI enables autonomous detection and resolution of software system issues.
• Redefining the future of software engineering — Agentic AI is poised to automate end-to-end software development and product lifecycles, driving significant speed and efficiency gains.
• Modernizing regulated industries with cloud and agentic AI — Agentic AI and cloud modernization are crucial for operational efficiency, resilience, and regulatory compliance across industries.
• Why agentic enterprises need to become learning systems — Agentic enterprises must capture operational knowledge to enable AI systems to learn and improve without constant model retraining.
• The OpenClaw-ification of AI — The AI industry is rapidly adopting agentic workflows, emphasizing persistent, scheduled, and multimodal AI capabilities.
• Agents are ready but your architecture probably isn't — AI value stems from outcome-driven design, robust data infrastructure, and proactive governance, not just model selection.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Amazon Science homepage.