The Safest Way I've Found to Let AI Coding Agents Run Real Infrastructure

YouTube · Watch/Read · (gathered 2026-09-10)

• The core challenge with AI coding agents interacting with real infrastructure is the necessity of granting them broad, risky access to credentials (e.g., cloud keys, database URLs), which poses a significant security threat.
• The solution proposed is to use an orchestrator like Kestra, an open-source platform, to wrap coding agents. This allows agents to invoke pre-defined workflows (written as YAML in Git) rather than having direct access to sensitive credentials, which are managed securely by the orchestrator within those workflows.
• This permission model ensures agents can only perform specific, controlled actions (e.g., restart services, scale replicas) on infrastructure, with full visibility into inputs, logs, and results via the Kestra dashboard, thereby mitigating the risk of unintended or malicious operations.