Muse Wonder: Coding and Reasoning for Long-Running Agents

Muse Wonder is now available on RouterHub as meta/muse-wonder. Currently in preview, Meta’s new model focuses on coding, tool use, and tasks that require sustained reasoning across many steps. For teams building software agents, it brings those capabilities together with a one-million-token context window and multimodal understanding.

The practical appeal is a model that can work through a sequence: understand the request, inspect relevant material, use tools, interpret the results, and continue toward a finished deliverable. Muse Wonder is especially relevant when the task extends beyond a single response or requires information from both code and visual material.

Coding that reaches beyond the first patch

Software engineering rarely ends at generating code. A useful agent must find the relevant files, understand dependencies, make a change, and respond to test failures without losing the original requirement. Meta positions Muse Wonder for this broader coding workflow, including complex repository work and tool-driven development.

Meta reports the following results for Muse Wonder at maximum reasoning effort:

Evaluation	Muse Wonder score
DeepSWE 1.1	78.8
SWE-Marathon	53.1
Terminal-Bench 4.0-AA	52.0

DeepSWE 1.1 measures software engineering across 113 tasks from 91 repositories in five languages, using pass@1. SWE-Marathon measures long-running software engineering through project tests; its published comparison uses different agent harnesses. Terminal-Bench 4.0-AA measures accuracy on tasks in isolated terminal environments, where executable checks grade the final state. These are Meta-published scores at maximum reasoning effort, not RouterHub-run measurements; outcomes depend on the evaluation setup and agent harness.

For development teams, a concrete starting point is a bug that spans a frontend component and its backend contract. Give the agent the reported behavior, the repository, and a failing test. A successful result includes a focused implementation, a passing regression test, and a clear explanation of the changed behavior.

Keeping long-running work on track

Long-running agents face a different challenge from chat assistants: intermediate results can change the next step. A search may reveal a missing dependency, a tool may fail, or a new instruction may narrow the scope. Meta highlights planning, subagent delegation, parallel tool calls, and recovery from errors as parts of Muse Wonder’s agent behavior.

That makes it relevant to work such as investigating an incident across logs, configuration files, and deployment history. The agent has to connect findings, distinguish a plausible explanation from a confirmed cause, and propose a change that matches the evidence.

The application still controls tool access, permissions, and how task state is stored. A large context window provides room for an extended task; it does not automatically add persistent memory between sessions. Teams should supply the context and records their workflow needs.

Reasoning across screenshots and documents

Muse Wonder’s multimodal understanding extends the inputs an agent can use. Meta highlights work with screenshots, charts, and documents alongside coding and reasoning. This matters when useful information is embedded in a visual interface or a structured page instead of plain text.

A support investigation, for example, might combine a customer’s screenshot with a written issue description and product documentation. A document workflow might require checking a chart against the surrounding explanation. The task is to connect the visible details to the question and produce an answer that can be checked against the original material.

For screenshot-heavy work, preserve the labels and surrounding interface context. For charts and documents, include the units, headings, and pages needed to interpret the evidence. Better inputs give the model a clearer basis for reasoning.

Muse Wonder on RouterHub

Detail	Muse Wonder
Provider	Meta
RouterHub model ID	meta/muse-wonder
Context window	1,048,576 tokens
Listed inputs	Text, images, files, audio, and video
Output	Text
Primary workloads	Coding, tool use, multi-step reasoning, and long-running agents

RouterHub’s model listing includes the input types above; confirm the format support in your chosen endpoint and integration.

Where Muse Wonder fits in your workflow

Start with work that combines several of the model’s strengths: a repository change informed by a screenshot, a debugging task that needs repeated tool calls, or a document analysis that must reconcile text with charts. These tasks bring coding, perception, and sustained reasoning into the same deliverable.

Use explicit completion criteria: the tests that must pass, the document questions that must be answered, or the evidence required to close an investigation. Record correct completion, human review effort, elapsed time, and tool retries across representative cases. Those results will show where Muse Wonder adds value to your workflow.