Muse Code Goes Generally Available with Cross-Session Collaboration and Programmable Agent Control via MSP

Meta’s Muse Code has exited beta. The update focuses not on one-off completions, but on enabling multiple long-running coding agents to coordinate their work. According to the [official announcements page](https://developers.meta.com/resources/blog/) and [release coverage](https://www.neowin.net/news/meta-graduates-muse-code-out-of-beta-with-new-features/), separate sessions can now message one another directly: if one agent’s changes affect another task, it can send a warning; when a blocker is resolved, the answer can also be passed to other sessions, reducing the need for people to manually carry context between terminals.

The new Workflow mode can orchestrate multiple sub-agents to handle complex engineering tasks. Setting reasoning effort to `ultra` increases the frequency with which workflows are launched automatically. The CLI also adds Rewind: pressing Esc twice lets users return to an earlier point in the conversation history, although they can select only points that the system considers safe, and confirmation is still required before completed work is undone. This design attempts to bring agent collaboration, history branching, and side-effect control into a single interface. Real-world deployments will still need to validate how conflicts are handled when sessions modify files concurrently, and whether “safe points” account for database, cloud, or external-tool operations beyond Git.

More significant for developers is `@muse-code/sdk`. Meta’s public [SDK repository](https://github.com/meta-models/muse-code-sdk) uses TypeScript, requires Node.js 20 or later, has no runtime dependencies, and includes type declarations. Clients control agent sessions through the Muse Session Protocol (MSP). The project includes protocol schemas, generated wire types, examples, test fixtures, and replayable compatibility transcripts. This enables teams to create, resume, and monitor Muse Code sessions from custom interfaces or automation systems instead of relying exclusively on the interactive CLI.

However, the SDK is explicitly labeled a pre-1.0 Developer Preview, and even minor releases may introduce breaking API changes. The public repository is also a mirror of upstream artifacts rather than the primary development location. Some tests require an actual `muse` host executable that is not included in the repository, so an independent checkout cannot perform end-to-end validation. Meta has also not published data on multi-agent workflow success rates, additional token consumption, the isolation model, or failure recovery. For now, engineering teams should pin package versions and build their own integration tests for approval flows, retry idempotency, write collisions, cancellation, and resumption after crashes before considering integration into production workflows.