Don’t Turn Back at the Edge (a Word of Encouragement for Agent Builders)

This post isn’t about tools or frameworks. It’s just a word of encouragement—for anyone out there building multi-agent systems and facing that moment where the work starts to feel impossible.

I’ve come to believe that’s the line where most people turn back.

When complexity increases, when innovation pushes into unfamiliar territory, it’s easy to start looking for an easier path. But that’s exactly where I tend to lean in. That edge—the place where most agent builders stop—is where I’ve found the most powerful breakthroughs.

When I’m building agents, if I can imagine them doing something, I assume it’s possible. And I don’t stop until I find a way to build it.

That doesn’t mean it’s quick, or easy. There are days where I spend hours or even weeks trying to solve a single challenge. And in those moments, a thought sometimes creeps in: “You don’t have to make the agent this smart.” Or, “Why not just buy an app that does this part?” Or even, “No one’s expecting this from an AI agent anyway.”

That thought usually shows up when I’ve been pushing the boundaries of what agents are expected to do—especially when I’m designing for intuition, scale, or seamless integration with human workflows.

Agents That Create Tangible Output

A recent example reminded me why I keep going.

I was building an AI agent that could generate documents—not just writing the content, but formatting it directly inside a Google Doc the way a human would. Most people today use generative AI tools like ChatGPT or Claude to get a piece of text, then copy and paste that into a separate document, format it, and send it off manually.

But what if the agent could do all of that for you? Not just write, but create the document, format the document in place—apply styles, add headings, space it properly, and email it on your behalf, all without being prompted over and over again. Not a text generator, but a thinking agent—one that remembers your formatting rules, understands your expectations, and delivers results directly into the tools you already use.

Is getting an agent to handle document creation easy? Not at all—especially if you’re doing it from scratch or have never built it before. But it’s possible. And that’s what I want to remind you of. It’s not just about prompting AI to generate text from your browser using ChatGPT. It’s about getting AI to create things that live in the real world.

So if you’re building agents that go beyond what’s expected, keep going. If you’re stretching their intelligence into unfamiliar territory—where things get complex, uncertain, or even a little uncomfortable—don’t retreat. That kind of work matters. It’s how we move from automation to true augmentation. And it’s where some of the most needed breakthroughs begin.

That kind of work also requires reflection. Look back when you need perspective. Study the systems that came before this wave of agentic design. Read the early research. Learn how others approached intelligence, collaboration, and system behavior—even before we called them agents. Then use that foundation to build forward. Let your imagination lead, but anchor it in context. That’s how we create what doesn’t exist yet—responsibly, clearly, and with lasting impact.

Thinking about implementing AI or multi-agent systems? I’d love to help or answer any questions you have. I also offer workshops and strategy support—learn more on my website!

Photo: “Woman Engineer Tech” by This_is_Engineering, released under the Pixabay Content License on January 19, 2024

When AI Agents Collaborate © 2025 by Nick Norman is licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International