Meta Launches Muse Code: A 24-Hour Coding Agent
Meta shipped a terminal coding agent built to keep going. Across a large repository, coordinating background agents, recovering after a crash, running for hours. Meta says one kernel-optimization run made more than a thousand tool calls over as long as 24 hours.
The number is the headline. It's not the story. When Oscar and I sat down for Episode 18, the thing I kept circling was everything around the model — persistent workers in isolated worktrees, an event log on your machine recording model calls, approvals, edits, and tool use. That's not a benchmark. That's an operating system for unattended work.
So here's the question we spent two hours on: what actually changes when an agent persists for hours, survives crashes, and coordinates other agents without waiting for you?
Signal or Noise
Oscar and I run every story through the same filter — is this signal you should act on, or noise dressed up as news? Here's how I called them this week.
Meta released Muse Code and Muse Spark 1.2
Beta on August 5, macOS and Linux, built for long jobs across large repositories. It launches persistent background agents, splits work across isolated worktrees, and leaves your working copy alone until changes are ready to review. I don't get excited about model-release week anymore, and I'd have skipped this if the pitch were "it scores higher." It isn't. The pitch is durability — an append-only local event log of approvals, model calls, edits, and tool use, so a crash means reconstruct rather than start over. Those are infrastructure decisions, not model decisions, and they're what decide whether you'd ever leave the thing running overnight.
Where I pushed back on Oscar: background agents multiply your bad assumptions as efficiently as your good ones. Parallel execution only pays when the result is easier to validate than doing the work yourself — otherwise you've traded four hours of writing for four hours of reviewing, and you've lost. Meta's 24-hour kernel-optimization demo shows the intended operating horizon. It doesn't prove reliability on a normal production repo. Signal.
OpenAI published ten Astra-generated results in math and theoretical computer science
Ten results across high-dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics — generated by an internal Astra, formalized in Lean, shipped with a 249-page paper, reasoning walkthroughs, and machine-checkable certificates. About $2,000 in tokens for the successful runs.
This is the least commercial story of the week and possibly the most important, because math is where you can't fake it, and because verifiable output beats a benchmark chart every time. But here's my problem: OpenAI chose the problems, ran the model, prepared the manuscripts, and announced the result. Lean checks that a proof follows from its assumptions — it doesn't check that the formal statement matches the question anyone cared about, or that the novelty survives expert review. The review cannot come from the same organization making the claim. Signal, pending outside review.
DeepSeek moved its updated V4 Flash API into public beta
Same 284B mixture-of-experts architecture as the preview, 13B active, but post-trained for agent work: 1M-token context, thinking and non-thinking modes, tool calls, OpenAI- and Anthropic-compatible APIs. Fourteen cents per million uncached input tokens, twenty-eight cents per million out, with setup guides for Claude Code and Copilot CLI already published. That combination makes it trivially testable inside a harness you already have.
Oscar's framing is route the cheap worker at search, triage, tests, and first-pass changes, then escalate the failures. Mine is a warning: low token prices hide retry cost. A model that needs three attempts is not cheaper than one that gets it right once, and almost nobody measures cost per accepted task. Test it on your repo before you believe the price tag. Signal.
The US frontier-model review framework reportedly excludes open models
Axios reported on August 4 that the White House finalized a voluntary pre-release review framework under the June 2 executive order — a classified cyber benchmark, and up to 30 days of government access to a covered model before it goes to other trusted partners. Covered models are reportedly closed-source, state of the art, and national-security relevant. Open models are out.
Two things worth holding onto. First, the framework text and the capability thresholds are not public; the open-model carve-out comes from reporting on a closed industry briefing, not a published document. Second, my actual position: open-model builders should not inherit a private review process designed for a handful of frontier labs, and the exemption is what keeps a voluntary test from hardening into licensing. Oscar's counter is fair — capability risk doesn't disappear because weights are downloadable, and a framework keyed to distribution format may miss exactly the systems you can't recall after release. I still have a lot more to say about this one. It's my unpopular opinion below. Signal, and the one I'd actually track.
Demis Hassabis left the DeepMind CEO role as Jeff Dean exited Google
Google's biggest AI leadership change since the Brain–DeepMind merger. Hassabis leaves the CEO seat but not the company — DeepMind chair, Alphabet chief scientist, still running Isomorphic Labs. Dean leaves after 27 years to start an AI-for-science company Google is investing in. CTO Koray Kavukcuoglu becomes SVP reporting straight to Pichai, with no replacement CEO named.
Oscar reads the split as leverage: move Hassabis out of day-to-day management and he gets range across research, chips, robotics, and drug discovery, with operating accountability landing closer to Pichai. I'm blunter. Executive titles are noise until they change model quality or shipping speed. Watch the next Gemini release and watch which researchers stay — that's the only evidence that counts. Signal, but not yet a product signal.
Stack Check
This week's rotating segment is Stack Check — each of us brings one workflow you can actually use now.
- Oscar: Codex and ChatGPT in one desktop app. His pick is about collapsing the gap between asking and doing. Fewer context switches, one surface.
- Matt: Warp as the terminal layer for Codex CLI and agent-heavy work. Mine follows from the Muse Code story. If agents are going to run long in a terminal, the terminal stops being a dumb window and starts being the place you supervise from. I want history, structure, and readable output when I come back to a session that's been running for an hour. That's the whole pick.
Oscar's hot take is the one I'd staple to the top of the episode: a coding agent is not ready for long, unattended work unless it can reconstruct every action after a crash. Durable state and an audit trail matter more than another benchmark win. I agree with him completely, which is rarer than it sounds.
My unpopular opinion: the review process is the tax
Everyone is calling this a light touch. Voluntary, relax. I have never seen a government program stay the size it started. Thresholds move. Thirty days becomes sixty. And the countries we are actually racing are not running this process and never will. We are building a very careful system for arriving second. Every day of review is a day of lead, and nobody is counting the bill.
And here is my warning. When they come for open weights, it will not look like lobbying. It will look like a safety report. Funded by a lab, written by people with real credentials, landing the week before a markup. Watch who pays for the study that says open models are too dangerous to release. That fight will be over before most builders notice it started.
Where I landed
Three things I'm taking out of this episode:
- Judge agents on their recovery story, not their benchmark. Ask what happens at hour six when the process dies. If the answer is "start over," it isn't ready for unattended work.
- The audit trail is the product. An event log of model calls, approvals, edits, and tool use is what makes an hours-long run reviewable instead of just impressive. Without it you don't have an agent, you have a very fast intern with no notes.
- Watch the funding behind the safety argument. Not because safety work is fake — because on open weights specifically, the money and the argument point the same direction too often for it to be a coincidence.
The honest summary of where Oscar and I ended up: the model got better, and that's not the interesting part. The interesting part is that the unit of work is now a session — long, unsupervised, and recoverable — and almost none of our habits are built for that yet.
Watch or listen to the full conversation below, and tell me which repository task you'd trust an agent to run for 24 hours.
Human in the Loop — Watch on YouTube · Listen on Spotify · Listen on Apple Podcasts
Every episode — full show notes, sources, and the archive — lives at podcast.vallyseed.com. The show is produced by VallySeed, where Oscar and I help organizations design, build, and deploy AI systems that create measurable competitive advantage.
