Meta’s Muse AI Routes to OpenAI Models via Azure

Meta's newly released personal AI agent, Muse, incorporates a multi-provider routing infrastructure that may utilize external foundation models for specific tasks, according to independent technical analysis of the software runtime environment. The investigation uncovered evidence suggesting the agent can route certain subagent sessions to OpenAI through Azure, alongside built-in support for Anthropic Claude models. Examination of Muse filesystem logs revealed a distinct model identifier labeled azure/muse-special. Log analysis of this session displayed signatures consistent with OpenAI Responses API conventions, including a gAAAAA encrypted payload prefix and call IDs formatted in OpenAI standard notation. This contrasts with Muse primary internal model, Avocado, which utilizes a distinct 32-character hex identifier and routes through Meta native inference infrastructure. The runtime model catalogue further confirms the presence of approximately fifteen Avocado versions alongside client implementations for Anthropic and OpenAI. Configuration files indicate that API credentials for these external providers are embedded within the agent daemon, accessible through an inference proxy service governed by a configurable kill-switch. Industry analysis suggests several operational rationales for this hybrid architecture. Meta may employ external models to handle tasks exceeding the current capabilities of its proprietary Avocado series, or to facilitate A-B testing for feature development. The setup also enables controlled experimentation with response routing and tool utilization without requiring user-side reconfiguration. Contrary to speculation that Meta might be distilling weights or reasoning data from frontier models, technical safeguards appear to prevent such practices. Analysis of the muse-special session demonstrates that raw chain-of-thought reasoning from external providers is encrypted at the source. The daemon is programmed to return only the final response and tool calls to the local environment, while the encrypted reasoning blobs are explicitly rejected by Meta reinforcement learning completion server. Consequently, external model interactions cannot be ingested for weight training. In contrast, conversations processed by Avocado are logged directly into transcripts and remain eligible for reinforcement learning optimization, subject to user opt-out preferences. The discovery highlights Meta strategic approach to personal AI agents, prioritizing a flexible inference layer that can dynamically allocate computational workloads. By maintaining routing flexibility across provider boundaries while enforcing strict data isolation for external reasoning, Meta appears focused on balancing performance with proprietary model development. The runtime environment design underscores a broader industry shift toward agent architectures that treat foundation models as interchangeable backend services rather than static frontends. As Muse continues its rollout, the underlying infrastructure suggests Meta intends to iterate rapidly on agent behavior through controlled external integration while keeping its core training pipeline insulated from third party data.