Meta’s new agent model family
Meta has introduced Muse Wonder and Muse Wonder Full, describing them as its most capable models so far for personal and professional agentic tasks. The announcement was published on 12 October and updated the following day. Wonder is available to developers through the Meta Model API and Muse Code, while Wonder Full is available to Muse Code subscribers. Meta says Wonder will later power more of the Muse product, beginning with subscribers.
The examples extend beyond a single chat response. Meta points to navigating an insurance claim, interpreting a lease, making sense of financial information and building a working app. These are tasks where the model must hold a goal, work across multiple pieces of evidence and know when to ask for clarification rather than guessing.
Active Memory carries relevant context
The Wonder family introduces what Meta calls Active Memory: the ability to retain information that matters across conversations and retrieve the relevant details for a later task. Persistent context can reduce repeated setup, especially when a project unfolds over several days. It also changes the privacy and product questions around a model because remembered information may influence future work long after the original exchange.
A useful memory system needs controls for inspection, correction and deletion. It should distinguish a durable preference from a temporary fact and avoid carrying information into an unrelated context. Organisations evaluating the API will want to know which memory behaviour belongs to the model, which is implemented by the application and how retention can be aligned with their own data policies.
Active Focus is about staying on task
Active Focus is Meta’s label for continuing towards a goal when work becomes difficult. The model is intended to try another approach, recover from errors and follow through, while checking with the user before consequential actions. That combination matters for agents: persistence without judgement can turn a small misunderstanding into a long sequence of unwanted actions.
Developers should treat the model’s persistence as one part of a larger control loop. Tool permissions, spending limits, timeouts and approval gates remain necessary. A model that says it will check before an important step still needs the application to enforce that boundary. Logs should show what it attempted, why a retry occurred and which result caused it to stop.
Expert domains require evidence and restraint
Meta highlights law, finance and health alongside frontier coding performance. Capability in those domains can help users organise documents and explore options, but it does not turn a general agent into a licensed professional. Applications should expose sources, state uncertainty and direct high-impact decisions to qualified human review. A lease summary or financial explanation is useful only when the user can trace the conclusion to the underlying document.
Coding presents a similar need for evidence. Generated software should be reviewed, tested and scanned for security issues before deployment. Muse Code may give Wonder a productive working environment, but the responsible unit of evaluation is the complete workflow: model, tools, repository permissions, tests and the human who accepts the change.
Safety for agents with personal access
Meta says the Wonder models were trained to avoid disclosing private information without permission, refuse assistance with harmful acts and continue supporting defensive cybersecurity work. The company applies its Meta Superintelligence Scaling Framework and plans a fuller report with additional safety evaluations. These commitments are particularly relevant when an agent can work across messages, calendars and documents.
The release signals Meta’s shift from a model that answers towards a model that continues. Developers can now assess Wonder through the Model API, and subscribers can use the stronger Full variant in Muse Code. The decisive questions will be how reliably memory remains scoped, how often focus turns into unproductive persistence and whether approval boundaries hold during real multi-step tasks.
Benchmarks should follow complete tasks
Agent models are best assessed with outcomes that unfold over time rather than isolated question answering. A useful evaluation might require Wonder to read several documents, propose a plan, use a tool, recover from a failed attempt and stop before an approval gate. Reviewers can then measure completion, factual support, unnecessary actions, recovery quality and the amount of human correction required.
Longitudinal tests are equally important for Active Memory. Evaluators should check whether the model recalls an approved preference weeks later, forgets a deleted fact and keeps separate projects from contaminating one another. These behaviours are difficult to summarise in a single leaderboard score, yet they will determine whether users trust an agent with continuing personal and professional work.
Developers should compare Wonder and Wonder Full using the same tasks, tools and approval policies. A stronger model may complete more work, but it may also take more expensive paths or attempt broader actions. Cost per accepted outcome, tool-call count and reviewer intervention provide a balanced picture. Meta’s API availability makes those workflow-level comparisons possible before a team commits sensitive production data or grants extensive permissions.