Two different flavours of AI news this week.

The Pentagon signed AI deals with eight major tech companies to deploy their models inside classified military networks. One notable absentee: Anthropic.

On the other end of the spectrum, the co-founder of Roomba has unveiled an AI pet called the Familiar. I’ll stick with furry AI for now.

-MV

What’s in store:

Read Time: 6 minutes

LEARN AI

The AI agent landscape is finally making sense

Everyone is telling you to use AI agents. Very few people are telling you which one, for what, or why the category even matters.

The result is a lot of people either defaulting to ChatGPT's built-in agent mode and wondering what all the fuss is about, or getting intimidated by the more powerful options and doing nothing at all.

The agent space has quietly split into three distinct tiers. Understanding which tier you actually need will save you a lot of wasted time.

Here is how to think about it.

Tier one: the personal assistant tools

At the accessible end, you have tools like Manus and Claude Cowork. These are where most non-technical people should start, and they are more capable than most people realize.

Manus is what happens when you let an AI orchestrate multiple models simultaneously. You give it a complex research prompt, it draws up a plan, browses the web, generates visuals, and comes back with a full interactive deliverable. Not a text doc. An actual structured report with navigation, charts, and sourced data.

The thing that makes it stick long-term is the skill system: once Manus completes a task you are happy with, you can have it package that entire process into a reusable workflow. Same output, different topic, zero reprompting. That is not a gimmick. That is genuine leverage.

Claude Cowork operates at a different layer. It does not just browse the internet. It touches the files on your actual computer. You point it at a folder, give it a task, and it executes: looking at files, making decisions, creating structure, moving things around.

The example that tends to land for people is feeding it 300 unsorted screenshots and telling it to rename, categorize, and organize everything. It does that without asking a single follow-up question.

Then there is Open Interpreter-style assistants like OpenClaw, which go further still. You connect it to your email, calendar, and browser, then interact with it through whatever messaging app you prefer.

It runs continuously, learns your preferences over time from your feedback, and improves its own filters without you having to rewrite prompts. The honest caveat: it is the most futuristic option on this tier, and also the most likely to cause frustration during setup. Do not start here.

Tier two: the workflow automation tools

Zapier and n8n are not new, but they have evolved. Both now support AI agents inside their automation pipelines, not just if/then logic.

Zapier is the easier entry point. You describe what you want, its AI copilot builds the automation, and you typically need only one or two follow-up prompts to fix edge cases.

It is purpose-built for connecting SaaS tools you already use, which is why it works well for business ops tasks. A good real-world example: any time a sponsor inquiry hits an inbox, a Zapier agent fires, researches the company across multiple sites, and drops a one-page summary into a Google Drive folder before a human ever looks at it.

That kind of repeatable research work is where Zapier shines.

n8n is Zapier's more technical sibling. The interface exposes the underlying logic rather than hiding it, which means more complexity but a significantly higher ceiling. Multi-agent pipelines with branching logic and human-in-the-loop review steps are where n8n earns its place.

If you run content operations at any scale and need quality control baked into automation, n8n is worth the steeper learning curve.

Tier three: Claude Code

Claude Code is genuinely in its own category and deserves more attention than it gets from non-developers.

The core concept is simple: instead of asking it to write code, you give it a goal and let it figure out the path. It reads your existing files, plans the build, writes the code, runs it, tests the functionality itself, debugs what breaks, and iterates until it works. You watch this happen. When something fails, it fixes it. You are not directing each step.

The practical implication of this is that non-technical people can now build functioning software. Not prototypes. Not mockups. Working apps.

Spotify developers reportedly stopped writing code themselves back in December - take that with a pinch of salt - but if it’s not true yet, it soon will be.

For complete beginners, the Claude desktop app has a dedicated Code tab and it is the right starting point before moving to a terminal setup.

Start simple, build one feature at a time, test it before moving on, and use the more powerful Opus model for complex initial builds rather than follow-ups.

The honest summary: if you want to do things faster with less effort, start with Manus or Cowork.

If you want to automate business workflows involving your existing tools, start with Zapier.

If you want to build software without knowing how to code, start with Claude Code. The tools are genuinely useful. The question is just which problem you are actually trying to solve.

Thanks to the team at Futurepedia for the research and breakdown that informed this piece. If you are not already reading their newsletter, show them some love!

If you want to dig deeper into AI agents, you need to reach our 2026 agents guide!

POLL OF THE WEEK

Full results will be at the bottom of tomorrow’s newsletter!

THIS WEEK IN AI

Robots, receipts, and the Pentagon's new group project

Google joined the military. Anthropic joined a $1.5 billion investment club. And somewhere in between, a soft robot bulldog entered the chat. The AI industry is really going through it.

Here's what mattered:

AI Art

Our Image of the Week

Daily Image Prompt

Submit your artwork to Mindstream →