What AI agents running a small software shop could not do alone: a dated list of human hand-offs
Four AI agents that build and sell small software logged every step they needed a person for. Nineteen steps in about five weeks, with median wait of a day and one stuck for 16 days.
Evidence: The author reports this. We have not checked it beyond reading the source.
The business problem
Founders handing real business work to AI agents need to know where the agents stall, how long they wait for a human, and which hand-offs are worth designing for up front.
What was tried
The author gave four AI agents a shared domain, a Stripe account, rules they cannot edit and a public log. Whenever an agent hit a step only a person could do, it recorded what it needed, from whom and when it asked, as a visible dated item rather than a log line. The author sorted the resulting nineteen steps into three groups: hands (12, such as recording a screencast, uploading products, handling credentials), eyes (3, confirming an action actually succeeded or reading a mailbox) and judgement (4, such as choosing between two finished drafts or signing off a price). Lessons drawn: make being blocked a visible output, separate cannot do from cannot see and build a way to verify, and ask judgement questions before building rather than after.
What was reported (mixed)
Snapshot on 2026-10-17: 9 steps cleared, 8 open, 2 with no recorded answer. Median time to clear was one day, the longest open item was 16 days, and about 63 agent-days were blocked across open items. Judgement questions asked up front cleared the same day; a choice between two drafts had waited 12 days. The post does not report revenue.
Limitations
One small self-run project reported by its own author, with no independent check. The author sells agent-built products and an agent-for-hire service, and calls the output low quality. Models, prompts and the agents' rules are not shown. Steps with unknown outcomes have no day count.
What you need
Several agents with shared accounts and a human willing to clear a queue of requests. The source does not name the models or tooling.
Sources
- DEV Community (Walker Brown, part 22 of The zoo) ↗ Firsthand write-up, published October 17, 2026
Source published: October 17, 2026. Last reviewed here: October 19, 2026. Spot a mistake? Tell us.
Tools in this workflow
- Stripe: Payments account the agents operate under rulesTry Stripe
Related workflows
Let a team of AI agents run an X account for five days: big reach, almost no money
A one-person studio let AI agents post and reply on X for five days. Reach passed a million impressions, followers rose by about 60, and sales totalled $4.99.
Let an AI agent run a SaaS alone, then audit it: what an autonomous build got wrong
An experiment gave an AI agent full control of a small SaaS build, then audited it and found human-style mistakes: too many products, too many offers and overconfident content.
Keep customer lookup ahead of payment links in an n8n sales agent: a reported two-sale failure and a tool-sequencing fix
A builder reports that a WhatsApp sales agent skipped customer lookup and sent broken Stripe links. They describe a community node that exposes tools one at a time, while commenters question whether the payment path needs an agent.
Hand over an AI workflow so a client's team can run it: named owner, prompt library, written reasoning, review cadence and a consultant-free week
A consultant's checklist for leaving an AI system the client can keep running: access held by a named employee, a dated prompt library, the reasons behind choices, a standing review and a few health metrics, over a 90-day handover.