NVIDIA’s new 'Self-evolving' AI wants to read your slack (for your own good)
Imagine an AI that actually learns how much you hate reading through 50 Slack notifications just to find one status update. NVIDIA is pitching exactly that with its latest open-source flex: a "self-evolving" agent stack designed to sift through your corporate clutter without leaking your company secrets to the public internet.
The setup uses something called Hermes Agent paired with NVIDIA’s NemoClaw blueprint. The big sell here isn't just that it can summarize your emails or GitHub issues, but that it "evolves" by learning your specific preferences and writing its own code to remember them.
I'd say it’s a very likely future for office work, provided your IT department doesn't have a collective heart attack at the thought of a bot reading your Outlook calendar. To prevent that exact scenario, NVIDIA is wrapping the whole thing in a security layer called OpenShell.
OpenShell acts like a digital bodyguard, enforcing a "network policy as code" that prevents the agent from talking to anything it shouldn't. It sits in a sandbox where it can see your Slack tokens but can't actually broadcast your internal drama to the outside world.
The "self-evolving" part sounds like the start of a sci-fi thriller, but it’s actually quite pragmatic. If you tell the bot you want a daily digest in a very specific, annoyingly pedantic format, the agent recognizes the pattern and writes a "SKILL.md" file to its own memory.
The next time you ask for an update, it doesn't have to guess—it just runs the skill it wrote for itself. NVIDIA claims these learned states can even survive a total system teardown and rebuild, which is a nice touch for anyone who has ever lost their custom settings in a software update.
But let’s be real: the success of this hinges entirely on whether users actually want to "teach" their AI like a new intern. NVIDIA’s example shows a user giving positive reinforcement triggers to save a format, which feels a bit like training a puppy to do your data entry.
It's a clever way to bypass the usual corporate PR fluff about "all-knowing" AI. Instead of promising a bot that knows everything out of the box, they’re giving you a tool that’s explicitly designed to be shaped by your specific, messy workflows.
Whether this actually saves time or just adds "AI Tutor" to your job description remains to be seen. If you're in the UK and want to see this stuff in the wild, NVIDIA is heading to the AI Summit London on June 10–11 to show off more of these live demos.
I'd say it's very likely: will this be the end of the unread notification badge, or just another "sandbox" for us to play in while we avoid actual work? Either way, the era of the self-writing office assistant is officially knocking on the door.
Sources: NVIDIA Technical Blog, NVIDIA Newsroom.
