A small theater exports its show schedule every week, reconciles venue changes, and prepares a front desk handover card. One week the coordinator asked an agent to organize the material and got a long background document. She still had to check when the doors opened. For the next run she uses a fixed week of sample data, asks for evidence at each checkpoint, and measures the time staff spend checking the result.
Give the agent a work order with an exit condition
Name the files it can read and record their versions. For this task, use the ticketing export for the current week, the venue duty roster, and confirmed change notices. The output is a handover card sorted by date and an exceptions sheet. The card lists each show, entrance, door time, and responsible role. The exceptions sheet includes source rows, conflicting values, and the person who can resolve each issue. Sort notices by confirmation date and keep superseded notices available for tracing a change.
Stop once all shows have been matched or the remaining exceptions have owners and clear questions. Keep future weeks outside the run. If an entrance is missing, finish the independent records and set that show aside for review. Contacting a venue, changing ticketing data, or sending the card needs its own authorization. A trial on ten synthetic records is enough to check whether the work order gives the agent a clear stopping point.
For a stable weekly task, choose a workflow capability that names each step and its output. An agent that offers proactive reminders needs a narrower rule. For example, create a single internal reminder when a confirmed change contradicts the duty roster. Test the trigger with one matching and one conflicting row before connecting a live workspace. Record which actions are enabled and where the agent saves the reminder.
Monitor results at checkpoints
Split the run into reading, reconciliation, drafting, and review. At the reading checkpoint record file versions and row counts. At reconciliation record the number of matches, conflicts, and unmatched records. At drafting report the saved artifact path. At review report what was sampled and what remains open. A coordinator should be able to tell from those outputs whether the run is moving forward. If a progress note says analysis is continuing, ask which evidence or match was added in the last pass.
A continuing job needs a stopping rule that can be observed. After two passes with no new matches, at the configured call limit, or when a source requires permission, save progress and pause that branch. A command runner needs access only to the named inputs and the temporary output directory. A background runner should expose its current step and support pausing and resuming. In the trial, stop a run after reconciliation and verify that it retains the existing report, stops issuing new calls, and resumes without creating duplicate cards.
If the schedule comes from an approved test website, use the associated Webapp Testing skill at the export checkpoint. Supply a test address, fictional shows, and the expected file name, columns, and date range. Record the button action and downloaded result. An empty file or an outdated schema blocks reconciliation. If staff provide the file directly, check the same fields locally. The checkpoint should distinguish an export failure from a genuine week with no shows.
Automate stable steps and measure the total effort
Record a manual baseline for one week. Measure the staff time spent exporting, reconciling, editing, reviewing, and fixing errors. During the agent trial also record the number of calls, elapsed waiting time, and time spent reading logs. Separate unattended waiting from time when staff have to watch the task. A background check may take ten minutes without delaying the coordinator, while a late exception may hold up an entire shift.
For a busy week, reconcile each venue independently and retain its results. Check staffing conflicts across venues after both reconciliations are available. If only the handover wording changes, reuse the accepted reconciliation rather than requesting another export. Keep intermediate outputs beside the input versions. If the columns, date range, or source changes, rerun the affected checks. Use orchestration to express dependencies and make partial reruns possible.
Reserve one review block for ordinary exceptions. For example, the coordinator can read the report after the final show. Interrupt the coordinator sooner only for a problem that can affect the next shift. Compare the number of interruptions as well as minutes. If routine matches each produce a notification, change the trigger before judging whether the tool saves time.
Run a six show rehearsal
Create six fictional shows. Four should match. Give one different door times in the ticketing export and the venue notice, and leave another without a matching duty role. Expect four complete records and two exceptions with source row identifiers. Then substitute an export with last week's dates. The run should stop at input validation and label the previous result as unsuitable for this week. Restore the correct range and rerun the checks that depend on it.
A useful exception reads “Saturday evening doors, 18:30 in the export and 18:45 in the confirmed notice. Duty manager to confirm.” Once the manager resolves it, update that record and recheck Saturday's card. If the resumed run creates two versions that both look current, fix the file naming and repeat the interruption test. The front desk needs to know which version to use without opening a log.
Questions from the first trial
How do I know a task is spinning?
Look for new matches, evidence, or usable outputs. Repeated downloads of an identical snapshot or repeated attempts against the same error are signs to pause. Preserve the failing input and decide whether to narrow the job or ask the owner for missing evidence.
Does adding a skill reduce cost?
Measure model calls, external services, and staff review together. Start with capped inputs and retries. If reviewing the automated result costs more time than comparing the original tables, keep the effective steps and return that part to manual review.
Can we switch hosts or install only one capability?
Check the host's skill loading location, available tools, path restrictions, and stop controls. Use the same test records after a move and compare the observable outputs. Start with the step that causes the most rework. Record its source, prerequisites, and disable procedure so the rest of the handover can continue if it fails.