A research update that knows what changed.

Make the first run a controlled comparison

Start in a Session with the supplied snapshots and ask for one saved result. The example has three sources and two cutoff dates. One source adds a private-preview capability, one current retrieval fails, and one remains unchanged. The expected output demonstrates the distinction before any recurring job consumes credits.

A baseline is a dated record of what was observed, not a permanent truth. Keep the source references and scope beside each claim. For real research, record page title, URL, retrieval time, and the relevant content so the next run can compare like with like rather than reconstructing yesterday’s evidence from memory.

Define material change before checking the sources

The example’s notification rule flags newly documented capabilities and access-policy changes. A newly documented weekly-export private preview qualifies, but it should retain the selected-account limitation. The Agent must not summarize it as a general launch for all users.

A useful rule distinguishes consequential changes from wording edits, navigation changes, and repeated announcements. Write down the threshold in the task instructions and test borderline examples. If every tiny page edit is treated as actionable, the resulting stream makes it harder to notice the changes the team actually needs to review.

Treat missing evidence as incomplete coverage

Harborlist’s current access guide was unavailable. That does not prove its access policy changed, disappeared, or remained the same. Keep the baseline statement as historical context and mark the current check incomplete. Report that only two of three current sources supplied usable content.

A failure rule can specify when to retry or when to alert a reviewer about persistent missing coverage. Keep that separate from the product-change rule. The fictional prompt asks only for a note about the failure; it does not authorize retries, alternate account access, or a real network action.

Save a brief that supports the next review

Request a change table, unchanged-source list, coverage register, and draft notification text. Each entry should cite both baseline and current source IDs when available. The unchanged Meadowdesk row matters because it proves the source was checked rather than silently omitted.

Keep the previous verified baseline until the new result has been reviewed. Do not replace an available historical source with an empty error response and call that the new baseline. When the review accepts a new observation, record which source and date became the reference for the next comparison.

Configure recurring execution only after the test works

Use a Schedule for a fixed time and a Loop for work that repeats after the previous run and a delay. Select the intended Project, Agent, goal, timing, and notification behavior in Workspace. Copying this template or opening a Session does not create that worker automatically.

Review the current model and usage rates before choosing the frequency. Loops use 10× model tokens, and computer usage is charged separately. Hosted workers can run in the cloud; a task that depends on a connected computer still needs that machine available. Enable unattended computer jobs only when both the task and Agent permit them.

Check the worker history before trusting silence

An absence of notifications can mean no material change, but it can also mean a failed run, inaccessible sources, or missing output. Inspect progress, job results, files, and run history. A useful recurring task records its coverage even when the notification rule says to stay quiet.

Pause or stop a worker when its topic is no longer useful, its sources change, or repeated failures need attention. Before rerunning a failed job, verify whether it already saved an output or delivered any notification. Reconcile that state first so a retry does not create duplicated briefs or repeated messages.

Try it with a small example.

An illustrative result.

Fictional sample, written to show the expected structure. Your result depends on your inputs and needs review.

Fictional expected run — 2026-09-10

Material change: Cedarboard now documents scheduled weekly export in private preview for selected accounts (B1 → C1). General availability is not established.
Unchanged: Meadowdesk still documents manual CSV export (B3 → C3).
Coverage: 2 of 3 current sources readable. Harborlist current status unknown because C2 has no retrieved page; preserve B2 as historical evidence.

Draft notification: “Cedarboard documents a selected-account preview for weekly CSV exports. Review eligibility before treating this as generally available.”
Baseline review: consider C1 and C3 after review; do not replace B2 with the failed C2 retrieval.

Before you use the result.

• The preview change preserves its selected-account limitation.
• The failed source is a coverage gap, not a product change or an all-clear.
• Coverage is exactly two usable current sources out of three.
• The unchanged source and retained historical baseline are documented.
• No worker, notification delivery, or automatic baseline replacement was created.