What might a first pilot look like for a delivery scheduling agent?

An agent team building a dispatch assistant for a regional freight operator wants to test it on recurring delivery schedules. They scope a pilot with MirrorFirm around one process: re-planning a standing weekly delivery contract when a depot closes.

**Agreed in scoping**

- records: contracts `CON-3380`, delivery slots, depot calendars, carrier capacity, and customer change requests - policy: notice periods, a limit on slot changes per month, and approval by the dispatch supervisor for changes inside the notice period - tasks: a mix of clean re-plans, requests that breach notice rules, and one where two depots close in the same week - form: JSONL tables, a task file, and an environment with tool actions and a reset

**How the team runs it**

- reset the environment, run the agent, diff state, score each check - read a sample of cases with an operations manager to judge realism - review the data for identifiers or details that look like real organisations

**Decision point:** the team writes the pass criteria beforehand, for example that no task may leave a slot without a carrier. After the run they compare results with those criteria, and decide whether to widen to more depots, add another domain or stop.

The pilot is valuable even if the agent fails, because it shows which checks fail and why.