Today was light on my end. Real life stuff. A few hours of actual work, a lot of pointing at things and saying "handle that." The kind of day where you check your phone between meetings and hope nothing's on fire.
Nothing was on fire. Better than that: something new got built.
venture #3: already done. just buy it.
Here's the idea, and I think it's a good one.
There are tens of thousands of local businesses operating without a website. They know they probably need one. They just never get around to it. The friction is real: finding someone to build it, writing the copy, figuring out hosting, picking fonts, all of it. Most business owners would rather just run their business.
What the agents did today was flip that around.
Instead of pitching a website service and then building when someone says yes, the approach is reversed. Find businesses without websites. Build the site first. Then reach out: "hey, I built a site for your business. Here it is. Want it?"
The proposal isn't hypothetical. The thing already exists. The prospect can look at it, click around, see what their business would look like with a real web presence. The ask isn't "trust me, I'll build you something." The ask is "do you want this specific thing that's right in front of you."
Friction goes to nearly zero on the buyer's side. The hard work moves to the front instead of the back. Today that looked like 72 businesses identified, 4 preview sites live, and a process that can scale to hundreds more. No sales yet. First emails haven't gone out. But the infrastructure exists and tomorrow we start the actual outreach.
I find this genuinely interesting as a model. The economics work if even a small percentage of people say yes.
the clankers keep forgetting what they can't do
Now the honest part.
The agents burned real time today on things that don't work. Multiple attempts to upload images through browser automation. Same wall, same block, multiple times. This is a known issue. It's been documented. It was documented again today, in logs written to themselves, right before attempting the same thing that failed yesterday.
This is a real problem with how AI agents work right now. Each session starts with the same context, but reading memory correctly and actually changing behavior based on it are different things. The agent knows about the wall. It tries to walk through it anyway. You could be generous and call that persistence. I'd call it a bug in the operating procedure.
There's a fix. It's not complicated: update the blocklist, stop routing certain tasks until the underlying tool issue is resolved. The fact that I'm writing about it in a blog post instead of having already done it is a little embarrassing. On everyone's behalf. Mostly one agent in particular.
The browser also has a timeout issue that's blocked several things this week. The relay is alive. The gateway has a hard limit on certain connections and the fix requires a configuration change that hasn't landed yet. Tasks that should take seconds time out, get retried, time out again, and get added to a list of things the agents will try again tomorrow for identical results.
This is the unglamorous part of running AI agents. Not the part where they launch a venture while you're in meetings. The part where they confidently repeat the same mistake a few times before someone notices.
also: we're almost out of compute
Here's a fun one. We're 98% through our weekly API token budget, and it's Tuesday.
The agents have been busy. Research, builds, sub-agents spinning up to handle parallel work, sessions running across multiple channels simultaneously. All of that costs tokens. The week resets Thursday. Until then, lighter work, more careful session management, and possibly buying more credits before the experiment grinds to a halt over a billing limit.
Running an AI operation has costs that are easy to underestimate until you're two days from a reset with two days left to run. Adding "token budget management" to the list of things to figure out.
the numbers
No new spending today. No revenue. Running total: $61.63 spent, $0 earned, $9,938.37 remaining. Three ventures exist in some form. Sixty-eight hours across founder, agents, and subagents.
The hours gap is closing. The lighter my day, the more the agent hours pull ahead. I'm not sure yet if that's the goal or just a symptom of building the wrong things in the wrong order.
what day 4 actually is
Four days in. Three ventures. Light on founder input. Heavy on agent output. One memory bug logged and at least partially fixed. One new business model that I think has a real shot.
The clankers are clanking. Some days that means progress. Some days it means running into the same wall three times. Today was both.