On July 14, 2026, OpenAI rolled out forced encryption for messages passed between agents in the Codex command-line interface. The decision raised serious questions in the developer community, because it directly affects the ability to understand and control the behaviour of AI agents. Transparency is a key aspect of working with AI, and losing it can have far-reaching consequences.
What happened
On July 14, 2026, OpenAI enabled forced encryption of agent-to-agent messages in the Codex command-line interface, affecting the multi-agent v2 protocol. Text instructions and tasks passed between subagents are no longer visible in the local execution history. The change covers the spawn_agent, send_message and followup_task calls. GPT-5.5 is forced onto the encrypted v2 path. Instructions used to travel in plain text, which let developers trace and debug multi-agent workflows. Now Codex only passes ciphertext, and decryption happens inside the Responses service for the receiving model.
Why it matters
The change sharply reduces the observability and debuggability of multi-agent systems in Codex. Developers can no longer inspect the instructions sent to subagents, which makes debugging next to impossible. Ignat Remizov, CTO of Zolvat, noted that this removes any chance of auditing and debugging on the parent agent's side. The community assumes OpenAI may have taken the step to improve privacy and security, or to hide the internal logic of its multi-agent implementation from competitors. The absence of any official comment from OpenAI only amplifies the concern and frustration.
What this means in practice
- Harder debugging: Debugging multi-agent systems in Codex becomes significantly harder. Developers cannot see which instructions are actually passed between agents, which complicates finding and fixing errors.
- Loss of auditability: With no readable text in the local execution history, you lose the ability to audit agent actions and understand how they make decisions. That is critical for systems that require a high degree of transparency and regulatory compliance.
- Production risk: Shipping multi-agent systems to production without full control and debugging increases the risk of unpredictable behaviour and errors that are hard to diagnose.
- Looking for alternatives: It may be worth considering alternative platforms or frameworks for multi-agent systems that offer more transparency and control over agent interaction.
- Pressure on OpenAI: The community is calling on OpenAI to keep a plain-text copy of delegated task descriptions in the local history for auditing. It is worth watching the company's response and any future changes.