Muse Code makes agent runs restartable

Meta’s Muse Code terminal agent appends every model call, tool run, approval, and edit to a local event log. Meta says that log lets a crashed run replay exactly and resume where it stopped; its background subagents persist through the session instead of being spawned for each task. Why it matters: Long-running agent work breaks when the record of what happened disappears with the process. A durable log turns recovery and review into something an operator can inspect instead of reconstructing the task from a fresh prompt.

Try this: For one coding-agent workflow, keep an append-only run record with planned actions, approvals, tool results, and verification output. Stop it mid-run, restart it, and check whether a reviewer can tell what remains before allowing another edit.