A few months ago I was doing the solo-developer dance: write a tool, zip it, write a sales page, upload it, update the README, check the site, fix the site, repeat. Every product launch was a two-hour manual ritual with a dozen places to drop something.
Today I run a 26-product digital goods store plus a content site with 3,000+ articles, and the only thing I touch manually is the occasional strategy decision. Everything else — generation, packaging, listing, monitoring, recovery — runs itself. Here's the stack that made it possible.
1. One factory script per product family
Every product is generated by a deterministic, re-runnable build script (Python, pure stdlib where possible — zero dependencies means zero pip install breakage in someone else's environment).
# build_my_tool.py — idempotent, re-runnable
# generates the tool, the README, the sample data, the zip, the cover art
# runs a built-in verification suite before emitting anything
Each builder runs its own validation before packaging: py_compile, real executions against sample data, exit-code contracts (0 = ok, 1 = something found, 2 = bad input). If verification fails, nothing gets emitted. The zip that reaches the store is the exact zip that passed the checks.
2. A cron factory does the listing
A nightly cron job takes the next item from a product queue and does the whole listing flow on Gumroad — via a Playwright session on a VPS that reuses a saved login. No API needed, no OAuth dance, no rate-limit anxiety.
The key detail: the listing script is built from a template of the previous successful listing. Every successful launch teaches the template one more edge case (this page has no Publish button, use the share page; wait 5s for the redirect; verify with a buyer-view curl afterward).
3. Verification is part of the pipeline, not an afterthought
After each listing, a buyer-view curl checks: HTTP 200, correct price, is_published=true, and the presence of the key descriptive keywords in the rendered page. If the buyer can't see it, the job is not done.
curl -s https://store.example.com/l/xyz | grep -c 'is_published.*true'
4. Monitoring + self-healing
- Watchdog crons check the health of the trading terminal, the VPN tunnel, and the content generator every few minutes. Failed = auto-restart or alert.
- The content generator produces ~30-40 articles/day; a periodic check verifies the article count is increasing and the sitemap matches.
- The email inbox is checked on a schedule for real orders vs. noise (verification codes, spam).
5. The "no-API desktop software" special case
Some of the most valuable automation targets have no API at all — a Windows-only trading terminal, a desktop chat app. The playbook there: UI automation, state-file parsing, watchdog/self-healing around the process. It's slower to build and it's a moat — most engineers won't touch it, which means the ones who can are in demand.
Why this matters if you hire engineers
The pipeline above isn't about selling tools — it's a proof that I can take a manual process and make it reliable by construction: deterministic builds, verification gates, exit-code contracts, monitoring, self-healing. That exact skill set is what any company building agentic workflows needs, whether it's AI agents for accounting, trading, or support.
If you're building something that needs automation, evaluation, and human-approval loops built in — or you just want the tools and the writeups — everything I've learned is at AgentChip.
The point isn't the products. The point is that a business can be a system — observable, verifiable, recoverable — instead of a pile of manual rituals.
Originally published on the AgentChip blog.
Top comments (0)