The same pattern shows up whenever agents move from toy workflows to systems with real state. As scope increases, weak guarantees create predictable symptoms: overconfident actions on stale data, brittle reasoning when meaning drifts, and compounding errors once the agent can write back.
The fix is to treat agents as what they are: systems that read, reason, and write against live operational data. That pushes you into establishing guarantees that most enterprise stacks provide only implicitly. Four matter more than the rest: freshness, semantics, safe write paths and lineage.
Freshness: Stop reasoning about yesterday
Many organizations have learned to live with staleness: batch pipelines, replica lag, caches, delayed CDC (change data capture), materialized views. Humans compensate with judgment. Agents compensate with confidence.
A common production failure mode is correct reasoning on the wrong time slice. The agent reads inventory that is minutes behind and triggers a reorder that collides with replenishment already in flight. Or the agent sees an incident marked “resolved” in one system while another system still shows the rollback pending, and it proceeds with a change that should have waited. Call these mistakes what they are: freshness bugs.
