OpenAI Codex now hides sub-agent instructions: handoff logs are gone

Since early June, users have reported that OpenAI changed OpenAI Codex’s experimental `multi_agent_v2` workflow so the prompts and messages passed from a main agent to its sub-agents are now encrypted. In practical terms, the plain-text handoff details that used to appear in session logs no longer show up there. You can still see the outcomes, but not the exact internal instructions behind a task split, a mistake, or a failed run.

Visual Studio Code Download

Those same reports say the change is required for the larger GPT-5.6 Sol and Terra variants. The lighter Luna model, at least for now, doesn’t appear to be affected.

If you use OpenAI Codex for more complex coding work, this matters. Once those readable inter-agent logs disappear, debugging gets harder. It can also create problems around audit trails, security, compliance, change management, and simple trust, especially for regulated teams, sensitive codebases, and customer-facing systems.

OpenAI hasn’t given a direct reason for the change beyond broad documentation about encryption at rest and in transit. So the guesses have filled the gap: model distillation, unauthorized API pooling, data harvesting, or an effort to protect proprietary orchestration methods.

You can access OpenAI Codex, including the experimental `multi_agent_v2` mode, through the Codex CLI in OpenAI’s developer tooling.