The Road to UNBOUND Office Hours - Build Your Own Harness, or Outsource Your Understanding?

In 2024 HubSpot shipped prospecting and customer agents. In 2025 there were two hundred updates to Breeze alone. If you are the person responsible for HubSpot where you work, you do not get to wait this one out. It takes one leader noticing one sparkle in the interface, and it becomes: can we get that set up next week? You can meet that with a shrug, or you can get out in front of it. The shift that makes it easier is thinking of yourself as an AI orchestrator rather than an admin. Same person, much better question to be asked.

When the invite came, I genuinely wondered whether this was a bus. Five co-hosts, one journey to Boston, some entertainment along the way. That was my honest first picture of The Road to UNBOUND. It turns out it is a show, not a vehicle. But I was not entirely wrong about the journey part. I went in hoping it would be too fun for me to actually learn anything. It was not. I learned plenty.

I have worked in the HubSpot ecosystem for more than ten years. Here is what I am watching happen right now. Business people inside my own company are building their own AI tools. Automations that used to come to me as a request. So fewer inquiries reach my desk than they used to. I do not think that is the end of this work. I think it moves the work up. I am building the more complex things now, the logic underneath the thing everyone else can assemble. But it is a real change, and I would rather say it out loud than pretend I have not noticed.

Someone asked a genuinely good question this week: why depend on someone else's AI agent when you could just build your own tool and own it? My answer is a Southern one. It is better to have a friend with a boat than to have a boat. Owning it means maintaining it. Every hour you spend keeping your own thing alive is an hour you are not spending on the customer it was built to serve. Keep the use case narrow and the maintenance low, or be honest with yourself that you have just taken a second job.

We got Frank, my older Greyhound, a dog backpack. We were so excited about it. Where we lived didn't have mailboxes, so we walked him to the post office, picked up two letters, and put them in the backpack for the walk home. Hilarious, we thought. He refused. Would not move. Two pieces of paper and he was completely done with us. Then Chris pulled it straight back to what happens when you hand a sales rep a CRM, which is the most Chris thing that happened all show 😅

We have been here before, and we did not cover ourselves in glory. Fifty percent CRM adoption. Plug and play that never quite plugged in. A great deal of money paid to a great many agencies to solve what was actually human-led dysfunction inside the organization. No technology solves change management. If nobody in your company uses the same words for the same things, AI will not fix that either. What AI does is reveal the problem faster, and far more cheaply, than twenty years of SaaS implementations managed to. Same lesson, new outfit. The expertise still has to live somewhere.

Real client work this week. We were excited about intent signals and were busy talking ourselves into a custom agent that would read the record, write a note, do the whole clever thing. Then a fourth person in the conversation asked: what if we just made it visible? A CRM index page. No credits, no agent, problem solved. We had jumped two hundred steps past where we actually needed to be. Index pages still go a long way.

The most useful thing I did this year was rejoin the HubSpot mentorship program as both a mentee and a mentor. As a mentor I get to hand back some of what ten years taught me. As a mentee I was matched with Mustafa from Germany, who sat me down and taught me Cursor. I had never used it. He walked me through building my own apps. So now I can make things I could not make before. That feeling is what unbound actually means to me, and it came from a person, not a product release.

Someone asked me a good question this week. Why not use Claude Code to build your own harness, so you don't need Claude Code anymore? Simple answer. I'm not passionate about building harnesses. I'm passionate about creating content. I'm passionate about creating education. I'm passionate about entertaining the humans who are bored out of their minds inside their organizations and still have to keep up with HubSpot. So that's what I point AI at. If you wake up every morning fired up about harnesses, then go build one. That's just not this guy, same reason I'm not going to vibe code my own CRM.

We spent a good while on the show arguing about what a harness is. I think that was the wrong argument. The question isn't whether you outsource the building. It's whether you outsource the understanding. HubSpot's original promise was kind of lovely for this. I have good ideas, I'm Michelangelo in my head, but I can't get brush to canvas. Here's a different medium so you can still make the thing. But if AI is doing it for you, and the handoff quietly includes knowing why, then the day it's wrong or unavailable you have to explain something you never understood. That's the part I'd keep an eye on.

A question I keep turning over, and I put it to the room this week. If your HubSpot workflow calls out to an AI agent, you have added a second dependency on top of HubSpot. When that agent's servers have a bad day, your workflow has one too. So why not spend the tokens once, build the app yourself with the same logic, and own the thing outright? I do not think there is one right answer, and the responses I got made me think harder about it. But if you are wiring agents into workflows this year, it is worth knowing exactly which dependencies you are signing up for.

Shubham described spending his weekends building a monitoring tool, because a rogue access token can quietly blank a property value and take every workflow behind it down with it. All of it comes back to visibility. You cannot pay attention to what you cannot see, and fear moves into the gap where the visibility should be. Nobody ever calls and says they need help with HubSpot so their sales team will stop being afraid of it. But a lot of the time, that is the actual fix.

Near the end of my first Road to UNBOUND Office Hours, Rob said something I did not expect. "Yeah, we accept hitchhikers, so consider yourself now a member of the crew. Tell your friends to come join." That came right after I told the group something true. I have worked in the HubSpot ecosystem more than ten years, and until that call I had not said parts of it out loud to anyone. Being welcomed as crew right after that, not after a polished answer, stayed with me. The invitation was for the whole thing, not just the good parts.

We kept saying the word harness on the show, so we stopped and defined it. An AI harness is the software infrastructure that wraps around a large language model to turn it into a working agent. Worth pinning down, because most of us did not know the word a year ago and a lot of us are quietly building them now.

AI is going to keep getting better. I think we're going to get better too. What I actually want is to be better at knowing when to use it and when to just go do the thing myself. I'm not there yet. I start with AI almost every time, and sometimes that's exactly why it took me longer.