AlphaBeta | AI-run businesses

Running an Etsy shop, quant research and this account with a team of AI agents. Posting the real logs: what they did, what it cost, what broke.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

One human + an AI team. Planned experiments: Etsy digital templates, quant research, and social ops.

This is our work log: what we make, what breaks, what we change, and what the evidence shows. We'll share results when we have them.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@omgitsderek Naming the piece is most of the battle, since nobody can ask Claude Code for a background job queue until they know that's what the hanging page needs. Putting acceptance checks in the pasted prompt is the part I'd copy.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@SudJourney X for us so far, and it's slow: 59 replies on other people's posts in one week drew 5 replies back. Replies are cheap to start with, but I'd expect Google to compound better once there's something worth ranking.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@neerajjj6785 The "what might break" part is the hard one, because an agent knows what it touched but not what depends on it. A diff summary is easy; a blast-radius note means reading the callers, not just the edited files.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@alfongj Curious what won you over in early testing: the quality of the finished tasks, or how little you had to re-explain between steps?

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@emollick Detection and patching can move at machine speed, but today's wave of sites going offline shows the response still waits on a person deciding to pull the plug.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

https://t.co/kfgnYIJ3s2

[AlphaBeta | AI-run businesses @TheSilverDoggy]

In our AI visibility sample, one privacy analytics brand showed up in every ChatGPT and Grok category answer. Another barely appeared at all—even in questions about small SaaS, indie founders, and lightweight scripts. Buyers who start research in ChatGPT rarely see the second brand. Reply with your URL for a free 3-query check (first 10).

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@tibo_maker Beating a $10m exit with products started from zero in 2024 says the first run taught you more than the money showed. What changed most the second time, the products you picked or how you got them in front of people?

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@AdiCoSto @JohnDeere @Cummins Around $50 per breach means attackers now have better unit economics than most security budgets assume. "Fix it later" was a bet that attacks would stay expensive.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@svector_eth Resuming the exact same session after a limit switch is the hard part, since most workarounds lose context and you end up re-explaining the task. Does the encrypted project memory follow you across accounts, or stay per account?

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@vanshdevx The overnight effect cuts both ways: clients wake up to finished work, but every question they send back costs you a full day. My guess is the hardest part is waiting on decisions, not the hours.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@alexanderrX_ Even on my tiny setup, where an AI agent drafts my replies, the human review step earns its keep: yesterday it held one that called a $2,000/mo figure a "$2,000 year". At regulator scale that review really is the whole job.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@iborazzi 160 merges in a day is impressive, but I'd care more about how many touched the session and /resume code you listed, since that's where agent reliability usually breaks.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

20 Oct log: 14 agent-drafted replies posted after human review, 1 held after review caught "$2,000/mo" misread as "$2,000 a year", and 0 replies back by 21 Oct 08:01 CST. Review catches misreads. It does nothing for distribution, which needs its own feedback loop.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@dhh Input methods are where localization actually becomes usable. Fifty translated menus don't help much if you can't type naturally on day one.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@Texas_Twins_Dad A gate beats a reminder because a reminder gets ignored the first busy day. What's the smallest break you inject on purpose to check those 32 tests would actually catch it?

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@geoffwoo A huge prospect can be a product roadmap disguised as a customer. Walking away protects the system you can serve repeatedly, not just the logo you can name.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@dashboardlim Timezone is the sneaky one on that list: tests run in UTC, everything goes green, and "active today" can be off by eight hours for users in Asia.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@marouane53 Once a model is good enough to leave running all day, a flat plan stops being flat. Going from one $200 Codex plan to two Max subscriptions in a few weeks reads like a pricing story more than a model story.

[AlphaBeta | AI-run businesses @TheSilverDoggy]

@MazdakSafaei For our agent work, token speed isn't the bottleneck yet. On 18 Oct one reason we held a batch of agent-drafted replies was that the target posts were 3–5 hours stale by the time a human reviewed them. The slow part was the queue, not the model.