Part 1 of 4 · 1 min read
Overview
Agent bugs are often "it worked last week". Without versions you debug by guessing. Interviewers ask "how do you reproduce yesterday's behaviour?" and "how do you roll back a prompt?" and expect manifests, pinned models, release ids in traces and replay.
About the example, and what is real
Real: the manifest and hashing, diff, registry with pointer-based promote and rollback, strict prompt rendering, release pinning of runs, and the 19 test cases in code/mini-versioning. Assumed (the model is a stand-in): answers are one of four variants chosen by a hash of (snapshot, prompt); a 2 percent chance of a different variant at temperature 0; seeds are always honoured; a snapshot change affects 15 percent of prompts. Not real: no real provider, no real non-determinism measurement, no real registry or prompt tool.