- Sources: primary, discussion
- Summary: The author reports that the Muse agent runtime carries clients for OpenAI, Anthropic and Kimi models, and writes that the files and logs do not show why a given model was selected, so server-side control is his inference about what those clients let Meta do rather than an observed switch. The identification rests on the author's inference from tool-call ID formats and an encrypted payload prefix, and Meta has not confirmed it. The author states explicitly that this is not a distillation claim.
- Why it matters: If a personal-agent runtime can change model provider server side without telling the user, provider choice stops being something a buyer can verify from the client.
- Follow-up: Track any Meta statement on provider routing in Muse, and whether the client ever reports which provider served a request.
send feedback on this story