ZeroNoise Logo zeronoise
Post
Coding Agents Alpha Tracker
Muse Code Makes Cost and Locality Part of the Coding-Agent Stack
5 hours ago •
4 min read
• 126 docs
Meta’s Muse Code beta and Muse Glimmer push cheaper terminal and local execution into the coding-agent stack. Practitioner tests and new orchestration tools point to task-level routing, reusable loops, and reviewable autonomy as the practical edge.

🔥 TOP SIGNAL

Meta is making cost and locality first-class coding-agent choices. Muse Code beta is a terminal agent for complete software-engineering tasks across large repos—planning changes, writing code, and validating results—powered by Muse Spark 1.2; Riley Brown places it between Opus and GPT-5.6 Tera and calculates roughly $5.50 in combined input/output cost for Meta’s top model versus $30 for Opus and $35 for GPT-5.6.

Muse Glimmer adds a 30B Apache 2.0 agentic model: Simon Willison generated an example with LM Studio’s 18.16GB build and says a 32GB+ machine leaves room for other applications. The practical shift is to measure completed-task cost, latency, and local execution alongside output quality—not simply choose the highest-ranked model.

⚡ TRY THIS

  • Route by total task cost, not token price. In DHH’s follow-on rewrite test, GPT-5.6 Sol High followed Fable’s plan: the first pass was 30% slower, but one follow-up reached parity at $43; DeepSeek V4 Flash could not get anything working despite many follow-ups. Kimi K3 was stopped after $60; when DHH let K3 Fast finish, it used $80 in tokens—about $55 at standard pricing—and was still slower and more expensive than Sol High. Start with a fixed acceptance test and a turn/dollar budget for the cheap model, escalate when it stalls, and log total task cost. Riley’s warning is the right accounting rule: low per-token pricing can be erased by extra turns.

  • Make optimization a separate, measured pass. DHH reports that a GPT Sol optimization run immediately made ttfx 63% faster—14× the original tte—followed by 16× and then 27× results in later rounds. After the baseline works, give the agent a bounded prompt such as: Profile this implementation against the original; optimize only measured bottlenecks; run the same benchmark; report before/after and regressions.

  • Orient before editing a neglected repo. ThePrimeagen opened a project untouched for more than a year, asked "yo ai, remind me of ....", and got up to speed in about three minutes. Use that first pass to recover architecture, entry points, tests, and unknowns; do not start implementation until the agent can restate the plan.

  • Use a reviewer agent to create a triage map. Kent C. Dodds asked Devin for a deep review of Kody. The returned review took six minutes and added 96 lines on a roughly 470,000-line codebase, citing zero TODO/FIXME/HACK comments, four uses of any, 606 test files, matching security invariants, and a “strong but not perfect” verdict. Treat this as a fast findings inventory for human verification, not as an autonomous approval gate.

📡 WHAT SHIPPED

  • Muse Glimmer open weights. Meta announced a 30B dense model trained for agentic use cases under Apache 2.0, with a stated 24GB-of-VRAM target; the GGUF build is available on Hugging Face.

  • Cross-harness coordination is productizing. Spotify launched Xirp, a vendor-neutral environment for managing sessions across Claude, Gemini CLI, and Codex; Spotify says 1,300+ engineers already use it and the service is now available to try. Riley Brown’s Buzz walkthrough shows the complementary team pattern: @mention Codex and Claude Code in one Slack-like thread, create a Cursor agent backed by Kimi K3, add it to channels, and watch which agent is working or requesting approval; the demo agent joined seven channels.

  • Kody v2026.08.10 adds built-in OAuth integrations. Operators can provision shared GitHub, Google, Slack, and similar OAuth apps through /admin/platform-integrations; Kent says onboarding feedback made this a necessary shift.

  • loop-library is now open source. Jason Zhou’s collection packages copyable prompts with loops he says deliver real-world results; he says most of the listed loops are already running in SuperDesignDev.

  • Durable agent chat got a concrete product treatment. Addy Osmani highlights Trigger.dev’s new chat agent, which survives refreshes, crashes, and redeploys and can pause for permission before a risky tool action.

🎬 GO DEEPER

  • Riley Brown — Meta’s NEW Muse Code is Here and Major Codex Updates: Watch the install → authentication → terminal muse → Wii Bowling flow. The important operational detail is the sandbox tradeoff: the demo configures YOLO mode to remove permission prompts and explicitly gives the agent full control, so reproduce that only in a disposable environment.
  • Repo to study — ttfx: Trace the progression from the first measured 63% optimization to 14×, 16×, and 27× claims. It is a compact example of separating agent-generated implementation from repeatable performance passes.

Editorial take: The high-alpha move is not picking one permanent model winner; it is making model routing, context recovery, optimization, review, and permissions explicit stages of one inspectable workflow.

Want personalized briefs on the topics you care about?

Create your own agent and get daily or weekly cited briefs on the topics you care about, shaped by the people, newsletters, podcasts, blogs, and channels you choose.