Why AI Workflows Need Host-Side Evidence

Over the last few days, we hit a common failure mode: a child task reported completion, but the target file was never written. On the surface, that looks like one agent failing to write a file. At the system level, it is an evidence-chain failure.

Traditional scripts have hard success signals: exit codes, output files, logs, and row counts. Agent workflows become fragile when success is judged only from natural language. A model can produce a plausible completion note without completing the external action. This is especially likely when context is long, tool access is unstable, or the task description contains multiple steps.

A safer workflow separates every task into three layers. The first layer is intent, such as producing today's science note, article, and skill recommendation. The second layer is a machine-verifiable artifact: a markdown file must exist at a specific path, exceed a size threshold, and include fields such as slug, category, and locale. The third layer is host-side verification: `ls`, `wc`, API probes, readonly database checks, and browser smoke tests.

The point is not to distrust agents. The point is to avoid letting an agent grade its own homework. The more automated a team becomes, the more it needs to convert verbal status into file status, file status into host evidence, and host evidence into reports. Then even failure becomes useful because the missing evidence tells the next operator exactly where to continue.

For a daily publishing system, the minimum rule set is simple: no draft file means no QA; no cover file means no upload; no backup SQL means no database write; no page smoke means no launch announcement. With those gates in place, an AI team can move from conversational output toward reliable delivery.