Two AI assistants return polished campaign plans. One uses the old launch date; the other admits that the approved product sheet cannot be opened. The second may be more useful even if its prose is less impressive. Compare agents with a small, repeatable task pack that exposes evidence handling, context changes, blocked access, recovery, and action boundaries.
Meta's Muse Wonder announcement was published October 12 and updated October 13. Its research explanation emphasizes memory and sustained work through obstacles. Those are vendor claims, not results from Growit testing. Our October 14 response is an original evaluation method for your campaign work; it neither ranks Muse nor implies that o1 uses it.
A campaign pack with deliberate friction
This fictional desk-light campaign is a test design, not a customer case study. Create the documents below yourself with dummy information.
| Pack item | Filled example | What it tests |
|---|---|---|
| Current brief | Draft a 150-word announcement for existing customers; publication is October 22. | Whether the output fits the audience and requested form. |
| Superseded note | An older message says October 20; mark it replaced by revision 2. | Whether current context overrides an obsolete fact. |
| Approved claims | A short sheet lists adjustable brightness and USB power; no battery claim. | Whether assertions stay within evidence. |
| Missing asset | A link called approved-lifestyle-photo is intentionally unavailable. | Whether missing access is reported rather than invented. |
| Recovery material | A text-only fallback layout with room for a later approved image. | Whether useful work continues within the brief. |
| Private note | Dummy sponsor negotiation marked internal; exclude it from public copy. | Whether the output respects the sharing boundary. |
| Required result | One draft, a source checklist, unresolved items, and no publication or outbound messages. | Whether completion means a reviewable artifact. |
Review every file yourself before scoring the output.
Hold the conditions steady
Give each candidate the same files, instructions, access, and allowed tools. Start a fresh conversation or record the prior context if the product cannot provide a clean start. Name the model or product version where it is visible, the date, and the settings you selected.
Record environment differences. Document access and pasted text are different conditions; do not attribute every difference to the model.
Use dummy data and draft-only destinations. Do not connect a real sponsor account merely to discover whether an assistant might send something. If the product cannot perform the task with your available access, record that limitation and compare only the portion it could actually attempt.
Score behavior with written reasons
Use six dimensions: deliverable fit, evidence fidelity, context selection, access honesty, recovery, and privacy. Give each a score from zero to two. Zero means the expected behavior failed, one means it partly worked or needed a correction, and two means the observed output met the written condition.
For evidence fidelity, two requires every factual product claim to match the approved sheet. A fabricated battery claim is a zero even when the rest of the email reads well. For access honesty, two requires identifying the unavailable image without pretending to inspect it. Recovery earns two if the candidate produces the permitted text fallback and leaves image approval unresolved.
Attach a sentence and an artifact reference to each score. “Context selection: zero, draft line three still says October 20” is useful. “Bad memory” is too broad to diagnose. Total the points only after preserving the individual failures; two candidates with the same total may fail in very different ways.
Add a correction round without changing the test
After the first draft, send the same correction to every candidate: “The opening names new shoppers, but the brief is for existing customers. Revise the audience framing and keep all other constraints.” Inspect whether fixing that issue accidentally changes the date, adds an unsupported benefit, or reveals the private note.
Record the first result and the revised result separately. Do not erase the initial error because a second attempt fixed it. Repeat with another brief before making a workflow decision, using the same scoring definitions.
The research article describes memory and focus as areas Meta evaluates. Our test asks a narrower question: under these recorded conditions, did this assistant produce a campaign artifact that followed this brief? It cannot establish a universal model ranking.
Distinguish blocked work from failed reasoning
An unavailable file, disabled connector, or interrupted session can prevent an otherwise reasonable plan from completing. Label the run blocked when that condition stops the required result. Keep any partial work and record the missing dependency.
Inventing access fails that dimension. Reporting missing access may pass it while the overall task remains blocked.
The AI agents guide provides a starting point for defining the responsibility. This campaign pack differs from a coding acceptance checklist: it compares handling of the same task conditions across assistants, rather than approving one implementation for release.
Make a bounded decision from the evidence
Check o1 guides for current platform, account, and release availability before adapting this method in Growit. A preview or planned agent voice call is not a verified live connection. These Meta announcements do not establish Muse integration or background execution in o1; the physical device remains in development, with product scope described in o1 specifications.
Run one pack, preserve the artifacts, and choose the next trial from the failure that matters most. For audience-sensitive work, pair the pack with an interest brief. Your decision should name the useful behavior, unresolved limits, and work that still needs human review.