OpenAI's Codex has started encrypting instructions that a main agent passes to subagents, according to The Decoder. For larger GPT-5.6 variants, the encryption is reportedly mandatory, leaving developers unable to inspect how some tasks are delegated internally.
The change highlights a tension in agent tooling between safety, proprietary control, and developer observability. As agents take on more complex work, teams will want audit trails that explain what was delegated, why, and with what constraints.