Balancing Speed and Safety: A Control Framework for AI Coding Agents
AWS, Thursday, July 30th, 2026
AWS proposes a control framework for AI coding agents that can open dozens of pull requests across repositories in an afternoon.
AI coding agents such as Kiro and Claude Code now generate features, tests, and refactors directly from natural-language prompts, and a single agent can open dozens of pull requests across repositories in a few hours.
That productivity comes with a trade-off, because agents optimize for task completion rather than for security or maintainability. AWS sets out a control framework covering identity and scoping for agent credentials, guardrails on what an agent may change, and review gates proportional to blast radius.
The post also addresses telemetry so security teams can see what agents actually did. The goal is to keep agent throughput without letting review capacity become the only remaining control.