You know how a company has different departments, one team does research, one does execution, one handles review? OpenAI just built that same structure into Codex, their AI coding agent.

Instead of one AI grinding through a task on its own, Codex can now spin up a whole crew of specialized AI agents, each with a specific job, all working in parallel, and then hand you back one combined answer when they are done. So if you ask it to review a piece of code, it can have one agent checking for security issues, another looking for bugs, another verifying the documentation, all at the same time.

You can also define your own custom agents with different AI models and instructions depending on the task. Want a lighter faster model doing the exploration work and a more powerful one doing the final review? You can set that up.

This is a big shift in how AI tools are being built. A year ago the conversation was about how good a single AI model could get. Now it is about how well a group of AI models can coordinate with each other. That is a pretty significant jump and Codex is one of the first mainstream tools to make it accessible.

Here is the thing nobody really talks about when they hype up how powerful AI agents are getting. The more autonomy you give an AI, the more you need to know it is not going off script.

NVIDIA just quietly dropped NemoClaw on GitHub, an open source tool that acts essentially as a secure cage for AI agents running on their Nemotron models. Every action the agent tries to take, reaching out to the internet, reading or writing files, making model calls, gets checked against a policy before it actually happens.

If the agent tries to reach somewhere it is not supposed to, it gets blocked and flagged for a human to approve.

It is early stage and rough around the edges by their own admission. But the fact that NVIDIA is building this kind of containment infrastructure now tells you something. The more capable these agents get, the more the guardrails around them start to matter. NemoClaw is not the flashy announcement of the week but it might be one of the more important things quietly being built right now.

The same day Codex Subagents went GA, a developer named Will dropped a free library of 425+ custom agent roles you can plug straight into Codex. Security experts, frontend specialists, data pipeline agents, debuggers, all pre-configured and ready to go. Curated from across GitHub.

If you are already using Codex this is worth bookmarking.