The one thing
01Meta 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. The line worth stopping on is further down: 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 not a feature, it is a position on what agents are for. A long task that loses everything when the process dies is not a long task, it is a demo.
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.
Shipped
05AI SDK 7.0.55 — batch APIs
Batch APIs land in the core ai package.6 If you have been hand-rolling queues around per-request calls to keep a bulk job affordable, that scaffolding is now upstream.
Claude Code 2.1.223
Owner wildcards — "owner/*" — now work in the strictKnownMarketplaces and blockedMarketplaces managed settings, so one entry allows or blocks every marketplace repo under a GitHub org.2
opencode v1.18.14
xAI login collapses to a single device-code flow that works headless, and structured mid-stream provider errors are no longer swallowed.4 Both fixes point at the same user: someone running an agent on a box they cannot open a browser on.
Cline 4.1.4 · CLI 3.0.50 · desktop 0.0.9
Skills now appear alongside workflows in the slash menu, and commands sharing a name are disambiguated instead of one silently shadowing the other.5 The desktop build is a single universal macOS download; existing installs migrate themselves.
Gemini CLI 0.54.0
Published at 01:35 UTC by the release bot.3 The notes are two links to older changelogs and nothing else. Diff the compare view before upgrading anything that matters — the release itself tells you nothing.
Promised, not shipped
Muse Code — beta, login required, no published price · Meta "new harness features and more powerful models" — no date
The conversation
01A 4B open-source model, post-trained with reinforcement learning, retrieved as accurately as GPT-5.6 Sol at roughly a hundredth of the cost. The post puts a typical multi-turn agentic search on Sol at over ten seconds and about three cents end to end.7
- @BedVibe_Studios · Hacker NewsReads the trend, not the claim
"This feels like the database equivalent of use the right data structure. We've spent two years assuming the biggest general-purpose model should do everything." Retrieval, reranking and generation each getting their own tuned model is the obvious shape once routing is cheap.8
- @breadislove · Hacker NewsThe pushback that lands
No standard retrieval benchmark, no reported metric. A hundredfold cost claim measured on a private evaluation is a marketing number until someone runs BrowseComp Plus or its equivalent.8
- @JCharante · Hacker NewsFirst-hand, unverified
Reports the same effect from their own testing — smaller models beating larger siblings on fact retrieval, apparently because the big ones overthink it. And asks the right question: why compare against Sol rather than Luna, which is cheaper and closer in class?8
The direction is right, and it is the arithmetic this desk led with in issue 001: capability per dollar is falling faster than capability, so the win is matching the model to the task rather than buying the largest one for everything.
The evidence is not. Castform sells post-training, the benchmark is its own, and the seller chose the comparison model — the pattern we refuse elsewhere in this issue. Take the shape of the argument, not the multiplier.
Community layer today is Hacker News only — the X and Reddit sweeps did not run. A narrower base than usual, stated rather than hidden.
Skip this
05Meta's own kernel benchmark inside the Muse Code announcement. A vendor measuring its model against a baseline it chose. We took the design details and left the numbers.
"Governments are making a dangerous bet on the AI boom." A leader column. Nothing in it changes what you ship this month.
Meta ran ads containing AI-generated abuse imagery. A serious story on the wrong beat. It belongs in your news reading, not a toolchain briefing.
Microsoft's AI revenue mostly traced to OpenAI. A disclosure for investors. No developer decision turns on it.
"AI will never become conscious." Philosophy without an artefact. There is nothing here to verify or act on.
Everything we saw
3636 candidates scanned · 11 used in this issue — the rest, with the reason each one was left out
End of feed. That is everything from the window worth your time.
Next issue tomorrow, 06:00 UTC — and if nothing ships, it will say so in two hundred words.