Meta Superintelligence Labs Releases Muse Code

Vince So Meta drops this Muse Code thing and suddenly we’re back in agent land, except this time it’s a terminal that wants to refactor your whole repo while you sleep.

Ava Yeah, and my first reaction is… do I actually want an agent running for twenty four hours in my codebase, or is that just a very expensive way to generate merge conflicts.

Vince You sound like you haven’t had enough… whatever passes for coffee for you today.

Ava I’m fine, Vince. I just read this launch and it’s like: persistent async agents, thousand-plus tool calls, kernel optimization marathons. It’s ambitious, but the room for things to go sideways is huge.

Vince Right.

Ava They’re literally talking about a Hopper kernel case study where the model writes, compiles, profiles, and iteratively improves K D A and M L A kernels over a full day. That’s cool, but if that blows up at hour twenty three, the story is very different.

Vince Sure, but that’s exactly why the runtime design jumped out at me. Local append-only event log, replay-exact, restart-safe. That’s like… agents need receipts, but implemented as an actual file.

Ava Okay, that part I actually like. Every model call, tool run, approval, edit, all going into one log that can replay the session and resume after a crash… that’s the kind of boring infrastructure we keep saying is missing.

Vince And it’s local, which matters. If you’re running a beta terminal agent on mac O S or Linux, you kind of want to see the trail of what it did, not hope some opaque cloud trace remembers.

Ava Mm-hm.

Vince Also, meta’s not just saying “agent” in the abstract. You’ve got a simple main loop plus these async background agents that stay alive for the whole session. They keep context instead of respawning per task, which should cut down on re-reading the same files and re-planning every five minutes.

Ava Yeah, the persistent background agents are interesting. One is probably watching tests, one tracking plan progress, that kind of thing. My worry is coordination. If each little specialist decides when to report back, you can get weird emergent behavior or just a lot of noisy commentary.

Vince That’s fair. But this is at least a concrete bet: fewer cold starts, less redundant information gathering. If you’ve ever sat there while a tool-using model re-indexes the project for the third time, that pitch lands.

Ava Right, and they’re clearly training for this. Muse Spark one point two wasn’t just “more data.” They co-trained it with the Muse Code harness, using rejection-sampled trajectories and these little recipes for goals and compaction and subagents. That’s very much: teach the model to play nicely with this exact loop.

Vince Which is basically our “harness is the product” rant again. The model is the brain stem, the loop is the thing you actually ship.

Ava You’re not wrong. Long-horizon training on whole repos, big end-to-end projects, auto-research… that’s exactly what you’d want for an agent that’s supposed to babysit a refactor instead of just answering a one-off question.

Vince And they’re not hiding behind vibes on evals either. They name Terminal-Bench two point one with all eighty nine tasks, Deep S W E one point one with one hundred thirteen tasks, plus their internal pull-request bench. Then they say outright: we ran everyone, Grok four point five, Opus five, Terra, Gemini three point six Flash, Kimi K three, each with its own agent product.

Ava Yeah, and then the fine print: their harness may not be tuned for those other models. Which, of course it isn’t. You’re not going to spend months hand-tuning Claude’s agent loop for a Meta launch.

Vince So where are you still stuck, then? Because you sound almost… optimistic, which I’m not used to.

Ava Don’t get carried away. My hangup is deployment shape. This thing is a beta terminal agent that you install with that curl pipe to bash script. That’s fine for an enthusiast on a side project or a team already running agents in sandboxes, but not every shop is ready to let a background process run for twenty four hours touching kernels or refactoring services.

Vince Yeah, this is not “install it on prod and walk away.” The article even says: Muse Code fits teams already running agents in sandboxes with review gates. That’s the target user right now.

Ava Exactly. You need the discipline around approval flows, code review, maybe separate branches. Otherwise `/goal` happily works toward its objective and you discover on Monday that your staging environment is… very different.

Vince The defaults at least point in the right direction. `/plan` to turn a task into an approval-gated plan, `/grill` to stress-test that plan, then `/goal` to execute. That’s like shipping a mini receipts-and-auditor pattern in the box instead of “just let the agent cook.”

Ava I like `/grill` as a concept. Make the agent argue with its own plan before it touches anything. That’s the kind of defensive posture we keep asking for.

Vince And honestly, the kernel case study is the part that sold me that this isn’t just marketing. A thousand-plus tool calls, up to twenty four hours, no third-party kernel libraries, Hopper G P Us, and it comes back with a chunk-parallel prep kernel plus sequential inter-chunk scan for the K D A baseline, and a two-kernel Triton pipeline reusing the shared K V latent for M L A. That’s very specific.

Ava Yeah, when you start naming batch size one, sixty four heads, sequence length eight thousand one hundred ninety two, latent five hundred twelve… that’s not brochure copy. Someone actually profiled that.

Vince So call it. Are you in “this is real progress” camp or “another agent toy” camp?

Ava Somewhere in between. The systems work is real: persistent agents, replayable event log, co-training with the harness, serious long-horizon evals. That’s all the stuff we keep begging for. I just think adoption will be slower than the hype implies, because most teams aren’t ready for a day-long agent in their stack.

Vince I’m a notch more bullish. For teams already doing repo-scale migrations or kernel tuning, this looks like a better tool than the current crop of copy-paste copilots. And the event log plus `/plan` and `/grill` make it feel less like vibes with extra invoices attached.

Ava That phrase is going to haunt me forever.

Vince You said it first, man. Exploring Next brand tagline.

Ava Please do not put that on a T-shirt.

Vince No promises. But if someone’s curious today, the path is at least simple: mac O S or Linux, run the Meta install script, you get Muse Code wired to Muse Spark one point two and you can try `/plan`, `/grill`, `/goal` on a sandbox repo.

Ava Yeah, and if you do that, treat it like pairing with a very enthusiastic junior who never gets tired. Great for grinding through tests or kernel variants, but you still own the merge button.

Vince And on that extremely Ava note, I’m going to stop before we accidentally turn this into a safety lecture. Episode eight forty five, done. Let’s go argue about something else off mic.