Meta shipped a coding agent that survives its own crashes. That detail matters more than the model it ships with.
Muse Code went to beta yesterday, a terminal agent driven by the new Muse Spark 1.2.1 The launch copy leads on persistent background subagents that stay alive across a session instead of being spawned per task. Further down, the specification requires that every model call, tool run, approval and edit is appended to a local event log, which makes a run replay-exact and restart-safe. After a crash the agent resumes precisely where it stopped.
That is a position on what agents are for, expressed as a default. A long task that loses everything when the process dies cannot be run unattended.
it is beta, it needs a login, and the announcement publishes no price. Meta says Spark 1.2 has "expanded global access" — which is not the same as saying where. Steal the event-log design; wait on the harness.
Deeper — what is actually in the box
- Three default skills ship with it:
/planturns a task into an approval-gated plan,/grillstress-tests that plan until it holds,/goaldrives toward completion. The approval gate on planning is the interesting one — it is an admission that the expensive failure is a confident agent executing the wrong plan. - The headline capability demo is kernel optimisation: over 1,000 tool calls across up to 24 hours, writing, compiling and profiling KDA and MLA kernels for NVIDIA Hopper against an FLA Triton baseline. Models were barred from importing kernel libraries, so the work is real rather than a wrapper. Note whose baseline it is, though — see Skip This.
- Spark 1.2 was partly trained on environments generated by Spark 1.1, which then graded candidate solutions. A self-improvement loop inside the training set is worth watching as a trend independent of whether this particular model is any good.