Codex is OpenAI's cloud-based coding agent, and it recently began encrypting the internal prompts passed between its sub-agents — the smaller, specialized agents it spins up to handle subtasks. This means developers can no longer read exactly what instructions one agent is handing off to another.

Sub-agent prompts act like internal memos in a multi-agent system. Encrypting them protects proprietary system design and prevents users from reverse-engineering or manipulating the pipeline. It also reduces the risk of prompt injection, where a malicious input tricks an agent by exploiting its instructions.

The tradeoff is transparency. Developers lose visibility into how Codex reasons across tasks, making it harder to debug or audit behavior. It's an early sign of a broader tension in agentic AI: as these systems grow more capable, the inner workings are increasingly treated as infrastructure to be secured, not code to be inspected.