In the deja build, an agent wrote its own gates. They all passed. One of them accepted any response containing an id field, even if nothing had been stored.
The prompt sounded careful. The test was weak. Broken code could ship green.
That changed what I spend time on. A prompt asks for behavior. A gate checks whether the behavior happened.
I still write instructions. I just trust the executable check more than the wording around it.
Every morning I ask my dog to stay away from the baby's high chair during breakfast. Every morning she gives in to temptation and chases a cheerio. I could keep asking. Or I could put the gate up.
One's a prompt. The other's a guardrail.
A typed language doesn't care what model you're using. A linter doesn't drift between versions. A health check that pings prod and expects a 200 .. that's been the same test since before I started programming. It'll be the same test after whatever model drops next week.
A prompt won't.
Models change. Instructions that help one run can become noise in another. The check stays legible: write the row, return the stored id, fail if either part is missing.
The note I left myself after deja was short: ask what bad implementation passes this gate. I reuse that question more than any prompt template.
Related
- deja .. the build where passing gates still hid a failure.
- gate-review .. checking what bad implementation can pass.