I gave my AI agents a massive upgrade this week

Hey crew,

I cover some more technical topics today than usual, but i’ve broken everything down in a really easy to grasp way.

Even if you're only using ChatGPT or Claude chat right now. It's worth the read. Understanding how agents work now makes the jump to using them way less intimidating later.

I’ve also got some cool skill to give away.

I sat down to write the newsletter this morning and I see OpenAI dropped GPT-5.4. Perfect timing. Sam Altman's calling it "the best model in the world, by far." I'm not to interested in the better benchmarks, but what the model can actually do now.

The headline feature is native computer use. Yeah, Claude and Gemini can already do this, but the difference is GPT-5.4 pulls ahead in specific areas. The model gets its own virtual computer, looks at the screen, decides what to click or type, does it, checks the result, and repeats. Nothing to install and no browser extension needed.

On tasks like navigating between apps, managing files, and completing multi-step workflows, it now outperforms Claude by a solid margin.

Where Claude still has the edge is in coding and more complex agent workflows, which is worth noting because OpenAI conveniently left out direct comparisons to Claude Opus 4.6 on those exact benchmarks.

They also put a particular focus on improving GPT-5.4's ability to create and edit spreadsheets, presentations, and documents (which is really meaningful upgrade). The kind of knowledge work most people actually use AI for day to day. They even shipped ChatGPT directly inside Excel and Google Sheets.

Other things worth knowing:

Google just dropped a single CLI that gives you access to every Google Workspace service: Gmail, Drive, Calendar, Sheets, Docs, the lot.

A CLI (command-line interface) lets you use apps and services by typing commands into your computer's terminal. The terminal is basically a text-based way to talk to your computer instead of clicking around with a mouse.

This matters because AI agents like OpenClaw and Claude Code also live in your terminal. So when a tool has a CLI, your AI agent can use it too. And the google workspace CLI was designed for agents from day one.

But I think the bigger story here isn't the tool itself, it's what it tells us about where agent tools are going. Justin Poehnelt, one of the creators, wrote a blog post making the case that CLIs are a better way for AI agents to use tools than MCP servers. And it's hard to disagree.

If you're sitting here thinking, what the f*ck is an MCP? All you need to know is it's just a way for your AI agent to use your apps like Stripe, Gmail, Notion.

MCPs are context-heavy. every tool call eats into your agent's context window, meaning your chats fill up faster and your agent starts forgetting things sooner. They're also slower to execute. CLIs are lean, fast, and already everywhere.

I use Claude Code for pretty much all my work, most of it non-technical. And I'm already opting for CLIs over MCPs whenever I can. If I want to give Claude Code a new tool like Google Workspace or Notion, I'm always going to look for the CLI first and only fall back to MCP if I can't find one.

This week I went on a bit of a skill-building spree. Two new Claude skills that have already changed how I work.

I also made a big infrastructure move this week. I migrated almost everything out of Notion and into Obsidian. Here's why...

Context is everything for your AI agents. You give them markdown files; your ideal customer profile, information about your business, brand voice docs, whatever they need to do good work.

The problem is, I've got about 20 agents built out across Claude Code and OpenClaw, and they ALL need the same context. That means duplicating the same files across 20 different agent workspaces.

And then when something changes in your business and you need to update those files? You're hunting through 20 different locations to find and update them all. It's a nightmare.

So a solution emerged: keep everything in one central place in Notion, and just point each agent there. But that meant every time your agent needed to understand something about you (which is every single chat) it has to hit the Notion API/ MCP to go find it. It's slow, and it eats up loads of context.

Obsidian solved it. It uses the same markdown format as Notion, but stores everything locally on your computer in your file system. There's an Obsidian CLI, and Claude Code and OpenClaw can traverse it and find what they need lightning fast and for way less context. One source of truth, accessible to every agent. I'll link a great beginner friendly video explaining this below in Brain Food.

I started using Claude Cowork this week. If you're still only using chat models like Claude or ChatGPT, Cowork is the step up. It's a proper AI agent.

Chat model = question → answer

Agent = goal → result

It's also extremely beginner-friendly. Not much setup needed to start delegating real tasks to it.

I ran a workshop this week for a mate's community of agency owners, showing them how to build an executive assistant in Cowork to handle all the boring repetitive business admin; emails, invoicing, scheduling, the stuff that eats your day without moving the needle.

Stay tuned because I'll be putting the full guide here in the newsletter over the next week or so. In the meantime, give Cowork a try.

Slim pickings this week, but this one's worth your attention.