How Meta’s New AI Agents Handle Complex Business Workflows

Meta Platforms Inc. released two new artificial intelligence models, Muse Wonder and Muse Wonder Full, designed to execute complex, multistep tasks for personal and professional use. According to a report by SiliconANGLE, Meta describes these as its most capable models yet for running autonomous agents. The company targets workflows that require sustained attention, such as processing insurance claims or navigating specialist fields like law, finance, and healthcare.

What new capabilities do these models offer?

The models introduce two distinct features aimed at improving reliability in long-running tasks. Active Memory allows the system to carry user context from one conversation to the next, surfacing relevant details when they become pertinent to the current work. Active Focus is designed to maintain performance during difficult segments of a task. When the model encounters an obstacle, it attempts alternative approaches and recovers from errors rather than stopping.

How does the model handle critical actions?

Meta states that the models check in with the user before executing important actions. This safeguard is intended to prevent unauthorized or incorrect changes during complex workflows. The system combines text, video, audio, and image inputs within a one-million-token context window. However, the output remains strictly textual. Meta trained these models at its Prometheus data center, utilizing pretraining and reinforcement learning methods previously introduced with the Muse Spark model in April.

How do they perform compared to competitors?

Meta’s model card compares Muse Wonder Full against OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Opus 5.5. In Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, Muse Wonder Full achieved a passing result on 13.1% of tasks, outperforming GPT-6 Astra’s 8.6%. The card notes that competing models were run on a best-effort basis and may not have been optimally tuned for the specific prompts used.

Where do the models lag behind rivals?

Performance varied significantly across different benchmarks. Muse Wonder Full led in DeepSWE 1.1 and SWE-Marathon, scoring 57.5% on the latter compared to Claude Opus 5.5’s 46.3%. However, Meta acknowledged the SWE-Marathon figures were not a controlled comparison. Claude Opus 5.5 surpassed Muse Wonder Full in ProgramBench, HealthBench Pro, and the OSWorld V2.1 computer-use benchmark. Codex and Claude Code generated the leaderboard scores for the rival models.

What does this mean for enterprise adoption?

The release signals a shift toward AI agents capable of handling end-to-end business processes rather than isolated queries. For organizations, the Active Focus feature’s ability to recover from errors reduces the need for constant human supervision during long tasks. However, the mixed benchmark results suggest that no single model currently dominates all professional domains. Leaders should evaluate these tools against specific workflow requirements rather than assuming universal superiority.

What to do and how to stay safe: Meta

• Review internal data policies to ensure AI agents do not expose sensitive information during multistep tasks.
• Test new models in isolated environments before deploying them for critical legal or financial workflows.
• Monitor vendor updates for performance changes and security patches related to agent memory features.
• Establish clear approval gates for any automated action that impacts external stakeholders or records.

Step-by-step guide: How AI Risk Management Works: The Hidden Failure Points

General security guidance from the Sandbox Dispatch newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the context window size for Muse Wonder?

The models support a one-million-token context window for mixed media inputs.

Does Muse Wonder Full outperform all competitors in coding?

No, it led in SWE-Marathon but trailed Claude Opus 5.5 in ProgramBench.

How does the model handle errors during a task?

It uses Active Focus to try other approaches and recover from errors.