This week I built an app where an AI agent decides what I buy. Specifically, my coffee.

Having an AI pick my vendor, or pick my food, was very interesting.

It's also a little nerve-wracking. The things you don't expect can be challenging.

I told the agent to keep adding coffee until the order hit $75 for free shipping. Shipping costs confused it, and it got stuck in a recursive loop because it couldn't hit the total I told it to hit.

I'm not the only one handing decisions to agents. Vercel says more than half of its deployments are now triggered by agents. When Jason Lemkin asked an agent where to host an app, it ranked Render first and Vercel fifth. [check: context from the brief]

But if you know what you want, a bit of randomness makes everything feel like a loot box. That's what's so exciting about it.

What made the coffee agent unique was that it worked almost like a smart contract: when the actions are done, the results are given. It's automated, but the reward still has some randomness.

You could set up rewards the same way in onboarding or other customer experiences. The reward is guaranteed, but you have to log in or check again to see what it is.

In behavioral psychology, a more randomized reward tends to get a better result, because there's some expectation and excitement around it.

Let agents make some decisions inside a walled garden. You get creativity and some uncertainty, in a good way.

Kyle Poyar found that 76% of AI-native companies now use variable or hybrid pricing, compared with 49% of SaaS companies. His advice: make variable pricing feel predictable.

I agree. Making it predictable removes a lot of the uncertainty around usage, especially with AI tools being so new.

It's hard for customers to know exactly what they're paying for.

Anthropic's new Claude Haiku 5.5 costs $0.10 per million input tokens, down from $1. Anthropic is also giving startups a free year of Claude Team.

Cheaper AI will help you automate more things. But the biggest issue right now is that people don't know what to automate.

They're automating things that aren't the bottleneck. You can always be more efficient.

The biggest problem is still knowing what to automate.

Salesforce is buying Listen Labs in a deal reported at about $2 billion. Listen Labs runs AI customer interviews and builds "digital twins" that simulate how buyers respond.

Right now, it would be hard to trust. But it's a starting point.

I can see a time when it's more trustworthy, because it's based on facts, not feelings: the data, the limits, almost perfect information. Less schmoozing, less emotional tie. Just a rational business decision.

Twin on twin, your digital ecosystem and theirs could match up to find the best pricing. That could lower a lot of friction, although getting there might take some time.

Twin on twin is a win-win.

You still need extensive testing in practice. Once you open an agent up to the internet, things can change.

Test it at every stage. Then give it access, and keep direct oversight while you test so you can pull the plug if needed.

Building Run Buy, I also learned to put guardrails in two places: physical (never give the agent your credit card) and in code (tell it it's read-only). Why both? If the agent knows it can't do something, it won't try.

I hate running, but I love coffee.

So I built Run Buy.

My runs sync from my COROS watch. When I hit my weekly 5K or 10K goal, I open a loot box that unlocks a new bag of coffee. The further past my goal I run, the fancier the coffee gets.

At the end of the week, an agent opens Chrome, adds everything to my cart on the coffee shop's site, and stops at checkout so I can approve it.

It's a fun way of building everything into a loot box. But most importantly, it's a way to motivate myself.

It also takes the purchasing decision off my plate. I often don't like making direct choices, especially when I want to try something new.

Sometimes I want a random wheel to help me.

Watch the Run Buy demo

Explore the public repository

I sat down with Johanna Johansson, who leads growth at Alguna, an AI-native revenue management platform.

She trained as a copywriter, moved to New York after watching Mad Men, and has spent 15 years at early-stage startups, often as a one-person go-to-market team.

Her point stuck with me: "If you want to be truly unique, you got to tell your own story in the only way that you can."

You still need to build your personal brand. For your content to really grow, there has to be personality in it, and the only way to get that is by sitting down, thinking it through and talking it through.

You can't prompt your way to being genuine.

We also talked about two things I keep coming back to:

Watch the full conversation

Last edition I said something new was coming. Here it is.

Juniply Challenge #001 is a weekend build challenge for people who wire up go-to-market. It's virtual and async, with a low barrier to entry.

I'm excited to get people building with new tools, building with AI, and letting their creativity shine. The brief goes out to everyone at the same time on November 13, and there will be cash prizes and swag.

Join the waitlist

Have you built a useful workflow, agent, prompt system, or strange AI experiment?

Send it to me.

It doesn't need to be polished. It does need to solve a recognizable problem or teach us something useful.

Future editions will feature one community build, including how it works, who should try it, its limitations, and what its creator learned.

I'm testing tools, building strange ideas, and learning what works in public.

Explore Juniply

Follow the 52 Weeks of Building challenge on YouTube

If you know someone who is also trying to understand where modern go-to-market is headed, forward them this edition.

If you're interested in building at all, even if you're brand new, join the hackathon.

At the end of the day, all AI is doing is allowing humans to be more human.

So make sure you're a good human.

Until next time,

From Brendan