# Codex begins encrypting sub-agent communications

Privacy enhancement creates auditability challenges in multi-agent systems

- **Source:** [Hacker News](https://news.ycombinator.com)
- **Published:** 2026-07-14
- **Topic:** [Software & Engineering](https://brifing.si/c/engineering)
- **Original article:** <https://github.com/openai/codex/issues/28058>
- **On Brifing:** https://brifing.si/s/47bac9ec-4fa2-48d9-954f-a98f0f048142

## Privacy vs Transparency

Codex's encryption of inter-agent messages in multi-agent systems complicates task auditing for users. This post-0.137.0 change creates tension between security improvements and operational visibility in complex workflows.

## Technical Implementation

Merge #26210 introduced encrypted_content field in InterAgentCommunication structure. Now spawn_agent and send_message operations encrypt payloads while leaving content field empty, significantly impacting debugging capabilities.

## Operational Impact

System administrators can no longer answer basic questions like 'What task was delegated to this sub-agent?' Rollout histories show only ciphertext, eliminating human-readable task descriptions from audit trails.

## Debugging Challenges

When sub-agent processes fail, root cause analysis becomes impossible. Encryption obscures which link in the task chain failed, dramatically increasing mean time to resolution for complex workflows.

## Expected Behavior

Users request maintaining structured, readable audit copies alongside encrypted delivery. Proposed solution: retain encrypted_content for model communication while adding separate audit_content field for human review in system metadata.

## Code Analysis

InterAgentCommunication::new_encrypted initializes content as empty string. The to_model_input_item conversion exclusively uses encrypted payloads, completely omitting human-readable content from persisted records.

## Related PRs

#26210 introduced the encryption change while #26753 addresses validation errors for encrypted tools. Unlike the latter, this issue concerns operational transparency erosion rather than schema validation.

## Workarounds

Developers might temporarily disable MultiAgentV2 or revert to pre-0.137.0 versions. However, this means sacrificing security enhancements for maintainability - an undesirable tradeoff.

## Long-term Solution

A hybrid mechanism preserving both encryption and auditability is needed. Messages should reach recipients encrypted while systems maintain readable versions in debug logs and metadata stores.

## System Design Lesson

This case study highlights how security measures impact operational visibility. Software architecture must carefully balance privacy requirements against the need for system observability and maintainability.

## Sources covering this story

- https://github.com/openai/codex/issues/28058 — 

---

_Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._