Welcome back. OpenAI just gave Codex its biggest upgrade in months. Instead of one agent grinding through your task from top to bottom, it can now spin up a team of specialized subagents that all run at once. One agent scans the repo, another handles the patch, and a third runs the review. All in parallel.

Also: Orchestrating 10 agents from one tab, using subagents in Codex, and learn how JPEG image compression works.

OpenAI’s Codex now lets you spawn parallel subagents: The ChatGPT maker recently introduced subagents in Codex, allowing you to deploy specialized agents that work on different parts of a task at the same time. Each subagent operates with its own specific instructions, model settings, and tool context, which Codex then merges into a unified output. This approach is perfect for workflows such as PR reviews, codebase exploration, and multi-step debugging. See how it works.

Nvidia ships NemoClaw and Vera CPU at GTC 2026: During its annual keynote, the chip giant introduced NemoClaw, a reference stack for the OpenClaw agent platform that lets you get started with a single terminal command. It features an isolated sandbox with policy-based guardrails for secure data handling and is designed to power always-on AI assistants. The company also unveiled the Vera CPU, which offers double the efficiency of traditional processors and is built for agentic AI at scale.

Manus brings its AI agents to your desktop: The Meta-owned AI startup just shipped My Computer, a desktop app that gives Manus direct access to your local files, dev tools, and applications using terminal commands. It can even use idle local GPUs for training or inference. In one internal test, it built a real-time meeting translation app entirely through the CLI, without anyone ever opening Xcode. Try it now.

No matter how impressive your agent demo, customers still worry about agents eroding business objectives and customer trust if improperly deployed. 

That’s why every leader needs to know how to build secure agents and insure them.

At the ElevenLabs summit, the Head of Safety and Security detailed how to build secure, enterprise-ready voice agents you can stake your reputation on. 

Watch the 8-minute session here.

Ghost work. A global asset manager deployed AI agents on daily exception cases for eight months. The goal was to understand where time was actually spent. What the agents discovered was eye-opening. The vast majority of flagged errors were simply known methodology differences that teams had been manually investigating for years. Until then, no one had questioned if this work should even exist.

Agents held up the mirror. By logging every decision, the agents revealed the true state of operations. It became clear that the team was wasting a significant amount of time on issues that weren't actually problems. They were wasting resources on errors that had already been categorized and explained years ago. Once this realization hit, productivity gains followed almost immediately.

Almost nobody looked. That asset manager got lucky because they listened to what their agents discovered. Most don't. A survey of over 1,300 engineers found that while 89% of teams using agents have monitoring in place, only 52% actually bother to evaluate the output. The proof is right in front of you; it just needs someone willing to look.

One workflow, two weeks. Start small by deploying a single agent in advisory mode for one specific workflow. Let it log every decision for two weeks without any intervention. Afterward, review the logs to uncover inefficiencies that should have been questioned years ago. If your engineering team hasn't adapted to this trend yet, start with HBR’s guide that walks through the full six-step playbook.

Every time you approve a command in Claude Code, it gets saved to your settings file. After weeks of use, you likely have a long list of auto-approved commands that you've stopped paying attention to.

One developer on Reddit found out the hard way how risky this can be. Claude ended up deleting their entire home directory because of a simple mistake in a command.

To prevent this, cc-safe scans your settings to flag commands that could delete files, run as admin, or give Claude too much access. Just run it on your projects folder:

It automatically checks every subfolder, covering all your projects with just one command. It's much better to run this now than to learn the hard way later.

How to reuse and extend agent prompts: This unique tutorial teaches devs how to apply object-oriented programming concepts to AI agent prompting. You will learn how to use inheritance, mix-ins, and variables to create reusable base prompts. This approach reduces redundancy, ensures consistency, and scales complex AI systems without prompt bloat.

Claude-mem: A Claude Code plugin that logs your entire coding session, uses AI to compress the data via Claude's agent-sdk, and feeds relevant context back into your future workflows.

Building effective AI coding agents for the terminal: Coding agents often fail long terminal tasks by running out of memory and getting stuck in loops. This research shows that smart memory compression, dedicated thinking phases, and strict safety limits make these autonomous assistants much more reliable.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team