Separate checkouts reduce file conflicts between agents. They do not automatically separate databases, network ports or credentials, and a session log does not undo external actions.
Launch features and current availability
Meta’s August 5 announcement introduced Muse Code beta with Muse Spark 1.2, persistent background agents and a local event log. Those are vendor-described capabilities, not PeopleAreGeek test results.
The current workflow documentation qualifies availability by build and staged rollout, and says the public Apple-silicon macOS artifact lacks the workflow engine. It asks users to specify isolated worktrees for parallel writers. This contradicts the old blanket description of every subagent automatically receiving one. Workflows and ordinary helper subagents are also distinct features.
The boundary a worktree actually creates
Git worktrees provide separate working directories and indexes while retaining shared repository state. That is useful for two agents editing different candidate changes. It is not an operating-system sandbox.
Our example has two independent checkouts whose tests both connect to one development database. Their source files remain separate, but a migration from one can change the schema seen by the other. The same issue appears when both servers try to listen on one port. Allocate separate test resources or serialize that shared step; do not infer isolation from the directory names.
A durable event log helps explain and resume work. It does not inherently make a retried external action safe or roll back an already-sent request. Review the patch and the commands with effects outside the checkout. This is an evaluation method, not a claim that we ran Muse Code.
Verify the tier as well as the token count
On September 8, Meta’s API pricing page lists Spark 1.2 input/output at $1.25/$4.25 per million tokens for Standard and $0.10/$0.20 for Contributor, excluding cached-input pricing. Standard prompts and completions are not used for model training; Contributor permits that use. These are API rates, not a flat price per completed coding task.
For an invented run with one million uncached input tokens and 200,000 output tokens, those rates yield $2.10 or $0.14 before other charges. Repeated tool output and extra agents can change the token totals. Assess that data-use choice against the material the session will actually send; a low token bill does not establish permission to contribute a client’s source code.
September 8: verify current workflow availability and API pricing, remove universal worktree isolation claims, and distinguish logs, sandboxing and external side effects.