Meta releases Muse Wonder models for developers and agent workflows

What changed

Meta presents the models as designed for personal and professional agent tasks, including work that continues across multiple steps. Its descriptions of memory, sustained task execution and performance are claims made by the company. This announcement does not establish a new Ads Manager feature or confirm that the models can operate an advertiser’s account.

The relevance for a marketing team is the expanding choice of systems for research and internal production workflows. PMC suggests starting with a bounded task: assembling a brief from supplied documents, checking a proposed plan against requirements, or building a small internal prototype. Specify the expected output and the evidence that would demonstrate it is correct.

Memory across tasks can be convenient, but teams should test whether information from an earlier brief is being applied appropriately. A detail that was correct for one client may be wrong for another. Keep project context explicit, avoid unnecessary client data in an evaluation, and check the product’s current controls before using it in shared work.

For work involving several actions, record where the model stops for human input and whether its proposed next step matches the brief. The useful comparison is not simply how much work it attempts. It is whether the result is accurate, traceable and easy for a colleague to review. Keep publication, campaign changes and spending outside an initial evaluation.

This is an AI/Productivity update with potential agency use, rather than a claim of proven advertising performance. Assess access in the specific product being used, and treat future Muse availability separately from the developer access announced now.