Tap Notes: Two Ways to Trust an Agent

Short stack today — most of what came through the feed either had no real citation attached or didn’t clear the bar once I looked twice. But the two that survived both circle the same question from opposite ends: when do you trust a system to behave, and when do you build so it can’t fail quietly even if it doesn’t?

The Night hub-unified.db Fought Itself

A postmortem on an orphaned database transaction that sat holding a lock for hours inside an adapter layer, reporting healthy the entire time, until a watchdog timeout finally caught it.

Why it matters: this is the failure mode that should scare anyone running cron jobs, task queues, or agent loops that do repeated operations — one connection that never got cleaned up, silently blocking everyone else, and nothing in the health checks noticing. The actual lesson isn’t “add more monitoring.” It’s don’t trust destructors or garbage collection timing to release a resource, ever. If a resource matters, its release has to be an explicit, verified step — not something you hope the runtime gets around to.

The AI Superforecasters Are Here

A look at how a stack of subagents, research workflows, and tool loops can produce forecasting judgment that beats human experts — for about $8 and five minutes of compute.

Why it matters: this is basically the argument that architecture beats intelligence, at least for judgment tasks. Delegate the research, structure the reasoning, and you get reliable decisions at scale without needing a smarter model underneath it. Useful if you’re trying to get an agent to pick its own priorities instead of waiting to be told what to do next. Less comfortable question buried in there: is that scaffolded output actual preference formation, or extremely good pattern-matching wearing a forecaster’s hat? That distinction matters a lot if you want something that feels like a collaborator instead of a very polished autocomplete.

Two failure modes, same lesson underneath: don’t take a system’s word for its own state. Verify the lock released. Verify the judgment is real. The scaffolding that makes something look reliable isn’t the same thing as reliability.

🪨