AI agent failures are often misattributed to model quality when the real culprit is infrastructure: misconfigured schedulers, shell alias assumptions, missing dependencies, and silent zero-exit failures. Drawing from 18 months operating 18 scheduled agents, the author argues teams must verify execution before evaluating output quality. Key lessons include distinguishing four execution states (ran with result, ran with no data, ran but couldn't access source, never ran), reporting 'unknown' instead of 'zero' when observability is incomplete, and extending standard telemetry with execution receipts — heartbeat signals proving a job actually fired, resolved its binary, and reached its data source.
Table of contents
About Suneet Malhotra485 Impressions1 Comment