Source: The Register — 2026-07-15
Summary
Codex CLI's multi-agent mode now encrypts the instructions a parent agent sends to its sub-agents before storing them, meaning developers can no longer read what a parent agent actually told a delegated sub-agent to do. The change became unavoidable when Codex CLI 0.144.4 shipped on July 14 and made the new MultiAgentV2 mode mandatory for GPT-5.6 Sol/Terra — though the underlying pull request had merged quietly back on June 5.
Key Takeaways
- Sub-agent delegation instructions are now client-side encrypted at rest, breaking the ability to grep or replay a local audit log of what one agent told another to do.
- The PR (#26210) merged June 5 but only became mandatory-in-practice on July 14, when CLI 0.144.4 forced MultiAgentV2 for the newest GPT-5.6 models — a slow-burn rollout, not a single announcement.
- The timing collides with the EU AI Act's Article 12 traceability requirements taking effect August 2, 2026, which push toward exactly the kind of inspectable agent-to-agent logs this change removes.
- A proposed developer workaround is a parallel plaintext audit log kept outside the encrypted channel — but that's a workaround bolted on after the fact, not a built-in guarantee.
Reel Script
Hook (18s / 40 words) Right as European regulators start requiring companies to prove what their AI agents told each other, OpenAI's Codex quietly made that impossible to check — it now encrypts what one agent tells another.
Core Concept (65s / 145 words) Codex's multi-agent mode lets a parent agent spin up sub-agents and delegate pieces of a task to them — write this file, run these tests, refactor that module. Until recently, developers could inspect what instructions the parent actually handed down, which matters enormously for debugging and for trust: if a sub-agent does something wrong, you want to know whether it misunderstood or whether the parent told it to. Now those delegation instructions get encrypted client-side before they're ever written to disk. Think of it like a manager whose memos to their team now get shredded the moment they're sent — you can still see the team acted, you just can't see what they were told to do. The change wasn't announced with fanfare; the underlying code merged back in June and only became unavoidable when the newest models made the new multi-agent mode mandatory.
Hands-On (55s / 120 words) Two concrete facts anchor this: pull request #26210 merged into Codex on June 5, and Codex CLI version 0.144.4, shipped July 14, is what actually forces every GPT-5.6 Sol and Terra user onto the new encrypted MultiAgentV2 mode — there's no opt-out at that model tier. The practical workaround developers are floating is running a second, plaintext audit log outside the encrypted channel, essentially duplicating the instruction trail themselves since the built-in one is now unreadable. And the regulatory backdrop makes this more than an inconvenience: the EU AI Act's Article 12 traceability rules, which require documented decision trails for AI systems, start applying August 2 — three weeks after this became mandatory.
Takeaway (25s / 52 words) If you're running Codex's multi-agent mode anywhere near an EU-regulated workflow, you need your own plaintext instruction log today, not a promise it'll be fixed later. Go check whether your compliance story actually survives an encrypted agent-to-agent channel.