Technical Context
I looked at the OpenAI Codex rust-v0.145.0 release, and there's real substance here: they haven't just tweaked the UX—they've stabilized opt-in multi-agent V2. For those building AI automation and dev pipelines, this isn't cosmetic; it's a change in the task execution model itself.
The key novelty I see is that sub-agents can now be assigned a separate model and reasoning level. So a parent agent can handle orchestration while I send child threads to different profiles—where I need a cheap sweep of the codebase or a heavier agent for complex logic.
This matters more than it sounds. Previously, model selection in many agent scenarios was too "session-based"; now routing gets closer to the specific thread and subtask level.
The second thing I immediately noticed is concurrency. Codex V2 normalizes parallel execution of multiple agents across a project: one digs into the repository structure, another checks dependencies, a third assembles a feature piece. Then it all comes back together into one response.
Also from the release: roles returned and agent navigation improved. In my experience, that's not a small thing—it makes a multi-agent system at least usable for daily work. When you can't quickly tell who's doing what and navigate the hierarchy, all the beauty of agent architecture quickly turns into chaos.
But I wouldn't romanticize it. Around V2, there are already questions about observability: the more autonomy sub-agents have, the sharper the need for auditing, debugging, and instruction control between agents.
What This Means for Business and Automation
For teams, I see three wins. First, it's cheaper to build mixed AI integration, where not every subtask runs through the most expensive reasoning profile. Second, large tasks that can honestly be parallelized go faster. Third, the architecture aligns more with a real engineering organization—an orchestrator and specialized executors.
Who loses? Those who hoped multi-agent alone would fix chaos in the process. Without routing rules, sandbox constraints, and clear observability, it breaks easily in production.
At Nahornyi AI Lab, we're precisely tackling these problems in practice: we don't just plug in a trendy agent mode; we build AI solutions for business so they are manageable, auditable, and don't burn the budget on flat ground. If you've got a scenario where a single assistant has hit a ceiling, I can work with you to hammer out the architecture and build AI automation tailored to your real process—no toy agents just for show.