Overboard
AI Agents for Solopreneurs: The Bottleneck Was Me
Most posts about AI agents talk about setup: hand an agent your marketing, content, or other work, and get your week back. I set up five agent departments on my Mac, each with its own goals and budget, and let them run unattended. The output wasn't bad. The biggest blocker was me.
What five AI agent departments did in eight days
Each department in Overboard has a written goal, a budget, the agents that do its work, and usually a captain: an agent that reviews recent runs and decides what runs next. Every run is a real Claude Code session on my Mac, on the Claude plan I already pay for. On October 13 I read through what had happened since October 5.
Department | What it did | Where it stalled |
|---|---|---|
Content Creation | 6 blog posts, 7 memes, 2 finished demo videos with posting copy | Review, edit and polishing made me not post them |
Design | 4 runs of social cards and favicons for three of my other products | Waiting on three one-word answers since October 7 |
Engineering | 1 task | Engineering work I still treat as sacred and not for overnight agents, so nothing on its list moved |
Marketing | Directory submissions nightly, social discovery daily, a daily log | Busy all week on the wrong things, these submissions did not bring traffic |
Research | Libraries downloaded, tested, reports created | No mission, lack of direction. Filled up GB of data testing github repos with no benefit |
Every agent did what it was told. The stalls came from how I had set up the work around them.
Everything ended in a draft, so I became the posting queue
After having problems with agents getting posts banned on HN and reddit with AI slop I obviously want to review what they post. Every run ends in a draft, and I publish. That was sensible when the agents were new. The problem is the backlog accumulates.
The two demo videos had sat finished for weeks. The Content Creation captain saw them in its recent runs, noticed nobody had posted them, and stopped making a third. That was the right call. It also meant a whole department had stopped because of me.
The Stark Insider piece on autonomous agents has a heading for it: "The bottleneck didn't disappear. It moved." For a solo founder there is no second reviewer to split it with. The fix I am building is a posting job. I approve a piece once and a venue once, and an overnight agent posts approved pieces to approved venues and reports the live links in the morning. Channels where a bad post is expensive stay mine.
Three one-word answers froze the Design department
Design ran four times, made social cards and favicons for three products, and then asked me to pick. Card A, B or C for one product, card A, B or C for another, and a cream or dark favicon for that second one. I didn't answer for six days, so four runs of finished work sat there and nobody could install it.
A question that needs one word should reach you wherever you are, with the files attached, and take one tap to answer. In Overboard a captain's question goes to Telegram or Slack with up to three of the files it is about. You tap an option or reply in the thread, and the next captain pass reads the answer.
The other half of the fix lives in the goal. "If I haven't picked within two days, use option A and tell me" is a sentence a captain can act on. Most design picks are reversible, so a default beats a week of waiting.
The healthiest department was busy on the wrong work
Marketing looked great on the board. It ran every night, logged every day, submitted directories, ran social discovery, and fixed a broken image on one of my sites. It also spent every night submitting Overboard to more directories (46 submitted, 12 live) when my October plan said to spread those submissions across six products.
The plan lived in a document the captain never read. The captain planned from its goal, and the goal still said "get Overboard listed." An agent optimizes what is on the page in front of it. When the scope changes, the goal the captain reads has to change with it.
How I'd set up AI agents for a one-person business now
One department per outcome, with a goal that names the scope ("six products", not "marketing").
A captain on every department that has a backlog, and plain schedules only for fixed jobs.
A budget per department in runs a day and minutes per run, plus a usage ceiling, so the night shift doesn't eat your day. I wrote up how I size a 24/7 schedule against Claude Max limits.
A written list of what agents may publish on their own, decided before the first night.
Questions that reach your phone with the files attached, and a default in the goal for when you don't answer.
A weekly look at each department as a whole. A run can look fine on its own while the department goes nowhere.
The jobs themselves are the easy part. If you want a starting list, I covered five marketing jobs to run overnight and what a nightly blog agent has to do alone.
Where the departments run
All of this runs in Overboard for Mac. There's no cloud: it turns your Mac into a 24/7 agent orchestrator on the Claude Max or Pro plan you already pay for, so there are no surprise token bills and nothing leaves the Mac. Departments hold the goals, budgets and to-do lists. A captain runs on a schedule, or whenever the Mac has sat idle for a stretch you choose, which is how mine plan while I'm away from the desk.
The work isn't only code. My departments write blog posts, browse social feeds, cut demo videos and Shorts, and make social cards. Every run is a real interactive Claude Code session launched with --permission-mode auto in Herdr, so you can open any of them and take over. The Mac has to stay logged in, and a keep-awake switch stops it sleeping while agents work. Videos, images and drafts attach to the run that made them. It's $49 once, with a 14-day free trial.