Build the boring plumbing first. Then add the AI.

We shipped The Hammer’s autonomous pipeline this week — cron jobs, Claude generation, Resend dispatch, the full machine. Then xAI deprecated their Live Search API the same morning Issue 2 was supposed to fire. The cron ran. It got an HTTP 410. The pipeline failed silently at 5 AM. Issue 2 still ships, and not because the AI worked — because everything around the AI did.

That’s the lesson worth lifting from this week, whether you’re building an AI workflow or buying one:

• The boring half is the load-bearing half. Resend kept dispatching. Supabase kept storing. Cloudinary kept serving. Vercel cron kept firing on time. The AI was the one piece that broke. Build the deterministic plumbing first, add the AI on top — reverse the order and an API outage means a dark page, not a degraded one.
• Wire the failure email before the success email. Our generate route emailed admin a “harvest failed” alert at 5:30 AM with the full error. The system told us what was wrong before any subscriber noticed. Adding monitoring after the first incident is too late — build it day one or you’re flying blind.
• Pin the SDK and the API version. xAI rotated the endpoint between when we wrote the integration (April 24) and when production hit it for real (May 7). Two weeks. Every AI provider integration is a 6-12 month maintenance contract whether you signed for one or not. Budget the time. Or pay someone to budget it for you.
• Have a manual fallback for every cron-driven thing. This issue reaches you because the cron is one of three paths through the same Resend pipe — manual SQL insert plus direct dispatch is path two; admin Send-It UI is path three. None require the others. If the AI brain has a bad morning, the body keeps the lights on.
• Don’t trust your own happy-path tests against APIs you didn’t write. Our parser tests were green. The HTTP integration was real-world stale within two weeks. Unit tests don’t catch wire-protocol drift. A weekly synthetic ping that actually calls the live endpoint would have caught this before launch morning.

The reason any of this matters: the system you’re reading this email through is the same one we’re building as a productized service for other founders. When we sell it, you’re buying the plumbing AND the lessons we already paid for. Including this week’s.

Start your intake →