My honest field notes on the verification gap no one's talking about, plus the complete guide + prompt kit that makes agent loops actually converge

A Claude Code plugin called Ralph Wiggum has been making the rounds in developer circles—named after the Simpsons kid who announces “I’m helping!” while contributing nothing useful.

The plugin does one thing: when the AI says it’s done, Ralph says “not yet” and makes it keep going.

We’ve spent years asking whether AI models are smart enough to do the work. We should have been asking a different question: who decides when the work is actually finished? And by what evidence?

The scene keeps playing out the same way. Someone gives an agent a real task—migrate a test suite, refactor a module, generate a report. The agent works, writing files and narrating its progress, printing a confident summary. Then it stops. The human checks, and the tests are red. Or the report is missing a section. Or worse: the agent never verified anything, but it sounded like it did.

The failure isn’t intelligence. It’s that nobody had standing to say “not yet.” We’ve been treating “done” as a conversational cue when it should be a contract. The next model won’t fix this, because “done” isn’t a capability problem—it’s an accountability problem. For most of AI’s recent history, accountability has defaulted to the model itself. Ralph inverts that.

The frame that makes this click: in agent land, the wrapper is the product. Model choice matters—but once you’re running agents on real tasks, most of the outcome delta comes from what happens when the model tries to stop. The checks, the loop, the verification layer—that’s where reliability lives. Teams buying better models while ignoring their harness are optimizing the wrong variable.

What’s inside:

• The metric that matters now: not first-pass success, but convergence

• What changes when agents move from drafting to doing (and why 2026 is the year it breaks)

• Where this shows up outside engineering—and why most teams will misapply it

• Where it fails—and the organizational collisions most teams haven’t seen coming

• The skill that’s becoming scarce: turning taste into checkable constraints