One Word, Four Meanings: How 'Replay' Works Differently Inside an AI Agent Runtime

A software engineering breakdown reveals that the term 'replay' carries four distinct meanings within an AI agent runtime codebase, ranging from a simple database read to a full, billable re-execution of a workflow. The cheapest form is an evidence replay, a GET request that reads run data across five database tables without triggering any model or tool calls. A second type handles reconnection after a dropped connection by replaying missed events and resubscribing to the stream. A third variant silently returns a cached result when a tool call arrives with a duplicate idempotency key, while the fourth actually re-runs the workflow and incurs new costs. The distinction matters practically because collapsing all four under one label can mislead users into treating an expensive re-execution as a harmless refresh.