Codex CLI: Multi-Agent V2 Reaches Stable
CodexCodex CLI's multi-agent v2 experience, previously opt-in and experimental, is now stable in version 0.145.0. The feature lets developers configure sub-agent models, reasoning levels, and concurrency, and restores agent roles with improved navigation between parallel agents. The move to stable status has also renewed developer discussion around transparency and debuggability of multi-agent runs, particularly around encrypted inter-agent prompts.
Featured Video
A video we selected to help illustrate this changelog
Key Takeaways
- Codex CLI's multi-agent v2 system moved from experimental/opt-in to stable in this release.
- Developers can now configure per-agent models, reasoning levels, and concurrency for parallel sub-agents.
- Agent roles were restored, along with improved navigation between concurrently running agents.
- InfoWorld reported developer pushback over reduced transparency from encrypted inter-agent prompts.
- A YouTube reviewer described a "multi-agent AI coding trap", cautioning that more agents don't automatically mean better results.
- The stabilization arrives alongside broader industry security scrutiny of AI coding agents, as highlighted by The Hacker News' coverage of malicious-code detection bypasses in agentic tools.
Sources & Mentions
5 external resources covering this update
The Multi-Agent AI Coding Trap: What Actually Works in Codex (2026)
Youtube
Codex Multi-Agent V2 update raises developer concerns over agent transparency
InfoWorld
Top AI Agents Built to Catch Malicious Code Can Be Tricked Into Running It
The Hacker News
Codex CLI 0.145.0 is out! Highlights thread
X (Twitter)
Release 0.145.0 · openai/codex
GitHub
From Experimental to Stable
Codex's multi-agent v2 system, which lets a root Codex thread spawn and coordinate multiple sub-agents working in parallel, has been evolving since it first shipped as an opt-in experiment earlier in 2026. With version 0.145.0, OpenAI marked multi-agent v2 as stable, signaling it is now ready for everyday use rather than being confined to early adopters.
Configurable Agents and Roles
The stabilized experience lets developers configure sub-agent models, reasoning levels, and concurrency independently per thread, so different sub-agents can use different models or effort levels depending on the task. Agent roles — restored in this release — and improved navigation between agents make it easier to keep track of what each parallel agent is doing inside the TUI.
Developer Concerns Over Transparency
The push toward stable multi-agent orchestration has reignited debate in the developer community about transparency and auditability. Coverage from InfoWorld noted developer concerns that encrypted prompts between sub-agents can make it harder to debug or audit exactly what instructions are being passed between agents, a tradeoff that comes with the added security of encrypted multi-agent communication.