Agents Come of Age? What's Actually Changing

We're one month into 2026 and already we see a few changes that set up the year in AI.

Anthropic - a company some doubted back in the summer - has been tearing it up in the coding world with Claude Code, and has now released co.work for everyone in the business space. Google has launched Personal Intelligence, connecting Gemini to your Gmail, Photos, and Calendar to deliver personalised responses. A clear move in a similar direction: AI agents that know you and works for you.

And then there's Clawdbot. An open-source personal AI assistant framework that has exploded across AI engineer Twitter, YouTube, and Reddit. Three weeks old. 80k GitHub stars. The community is split: some calling it the best thing ever, others calling it a security nightmare. (It's now called Moltbot after a cease and desist from Anthropic - make of that what you will.)

Something real is shifting. After building two different personal AI agents over the past month, I think I finally see what.

2025: The Year Agents Were Spoken Of But Not Seen

Remember when 2025 was supposed to be the "Year of Agents"?

For coding, it absolutely was. Claude Code, GitHub Copilot, Cursor, and other coding assistants delivered genuine productivity gains. Developers were faster and more efficient.

For everyone else? It was a hype train most saw little evidence of.

Most agents couldn't remember what you said yesterday. Didn't know what day it was. Wouldn't alert you to problems. Lacked domain-specific capabilities. Lost the thread on anything spanning multiple sessions.

For specific tasks, those who used them got huge value. Yet the data tells a different story about broader adoption: Gallup reports that nearly half of U.S. workers still never use AI at work. Only 30% of managers use it frequently. McKinsey found just 23% of organisations scaling any kind of agentic AI, and no more than 10% in any given business function. The agents handled tasks, not jobs. That's why they underdelivered on the promise.

My Journey: Cassie and R2

Over Christmas 2025, I built Cassie - my AI Chief of Staff. Connected it to my email (read only), CRM, and calendar. Gave it shared ops responsibility. I gave it long-term memory and and the ability to learn - it could see everything but couldn't send on my behalf. All running on our secure platform at TAU.

Cassie quickly became the best tool I'd ever used for taking notes, tracking follow-ups, managing projects. As someone who can struggle with the admin side of running a business, I already couldn't do without it.

Then just a few days ago, I fought through a messy Windows installation (lost that battle, switched to Linux) and got Clawdbot live on a dedicated Mini PC on my desk. Meet R2: my builder and doer.

Here's the key insight: Cassie and R2 do different things. And that's by design.

The Access Principle

Think about how companies actually work. You don't give every employee access to the finance folder and HR records. Not because you don't trust them, but because they don't need it - and limiting scope lets them focus on their actual job without worrying about accidentally touching something they shouldn't.

The same logic applies to AI agents.

Cassie has access to my email, CRM, and calendar. It handles operations and communications. R2 doesn't have that access. It sits airgapped on its own Mini PC, with only the files I explicitly give it and access to the wider web. That constraint is a feature, not a limitation. It means I can dial up its capability - giving it more dangerous tools for building, coding, and creating - precisely because its blast radius is contained.

In just a couple of days, R2 has been doing research, generating images, writing slide decks, and coding prototypes. It's more capable because its access is more constrained. I can't wait to train it on more skills and give it more tasks. The vision of talking to your friendly AI helper, going to bed and waking up with work done is starting to become a reality.

Two agents. Different roles. Security through separation. This is the pattern that matters.

What's Actually Different Now

We've been thinking about AI too narrowly. Judging early versions for what they were, not what the technology can become.

These systems are still very much human-led. R2 is useless without me - I'm setting direction, reviewing output, making the calls. Yet it's also proactive, and able to do things I can't. Either more quickly, or more capably, or both. I can't draw for toffee. (For my American readers: I can't draw worth a damn.) R2 can generate visuals in seconds.

This isn't AI replacing people. It's people superpowered by their AI assistants.

Something genuinely new is emerging. People will figure out the problems - they always do. This is genuinely powerful.

What's Next

I think we'll see a trend of more executives having their own personal AI. There could be a cottage industry in setting up Clawdbot/Moltbot servers in the near term. Longer term, it will be an enterprise feature - TAU already has a prototype.

The agents will add more skills: finance, travel booking, media planning. Whatever the job requires.

Many will worry about the human element. And it's right that we consider it. Yet I'm more convinced than ever that the better these tools get, the more they multiply people rather than replace them outright.

And such a productivity boost, I as an optimist believe, will be a good thing for society.