Meta’s Muse Code beta brings long-running agents to the terminal

Meta has just rolled out Muse Code in beta, bringing persistent background agents to tackle large codebases from the terminal. Powered by Muse Spark 1.2, it supports parallel sub-agents and a local event log to resume work after crashes.

Meta’s Muse Code beta brings long-running agents to the terminal

TL;DR

  • Meta released Muse Code (beta), a terminal coding agent for large repositories powered by Muse Spark 1.2
  • Long-running background agents: Persist across sessions, accumulate context, and split work via parallel sub-agents in isolated worktrees
  • Local event log: Records model calls, tool runs, and edits; enables crash recovery and resuming without re-prompting
  • Benchmarks: Muse Spark 1.2 scored 82.9% (Terminal-Bench 2.1) and 59.3% (DeepSWE 1.1), below Opus 5
  • Internal bench and speedup chart: 70.6% Meta Internal; +68.7% speedup metric; internal benchmark comparisons drew X criticism
  • Access and licensing: One-command install; includes a contributor tier; pricing and open-source status not specified; more details “soon”

Meta has released Muse Code in beta as a terminal coding agent for large software repositories. Mark Zuckerberg describes it as a tool for planning changes, writing code, and validating results, powered by the coding-focused Muse Spark 1.2 model.

Meta releases Muse Code beta for long-running coding tasks

Muse Code is designed around specialized background agents that remain active throughout a session and accumulate context across tasks, according to Zuckerberg. Larger jobs can be divided among sub-agents working in parallel in isolated worktrees, leaving the main working copy untouched.

Zuckerberg claims the setup built six game features simultaneously during testing without collisions. The post does not provide details about the test project or how the results were evaluated.

The system also records every model call, tool run, and edit in a local event log before execution. If a task crashes, Muse Code can resume from that log without losing work or requiring another prompt, Zuckerberg states.

Muse Spark 1.2 benchmark results

Meta’s published comparison lists Muse Spark 1.2 behind Opus 5 and GPT 5.6 Terra on Terminal-Bench 2.1 and DeepSWE 1.1:

ModelTerminal-Bench 2.1DeepSWE 1.1
Opus 586.7%65.0%
GPT 5.6 Terra81.8%64.8%
Muse Spark 1.282.9%59.3%
Grok 4.581.6%56.6%
Muse Spark 1.176.2%53.0%
Gemini 3.6 Flash78.9%40.0%

On the Meta Internal Coding Bench, Muse Spark 1.2 scored 70.6%, compared with 79.4% for Opus 5 and 65.4% for GPT 5.6 Terra. That internal result drew criticism on X, with commenters questioning the use of an in-house benchmark and asking why GPT 5.6 Sol and other models were absent from some comparisons.

A separate speedup chart measures cumulative tool-call count against speedup versus baseline. It lists Muse Spark 1.2 at +68.7%, below Opus 5 at +74.0% and GPT 5.6 Sol at +71.2%, but above GPT 5.6 Terra at +65.1% and Gemini 3.6 Flash at +62.5%. These figures describe the supplied chart’s metric rather than overall coding performance.

Zuckerberg also highlights a longer-running test in which Muse Spark 1.2 made more than 1,000 tool calls over 24 hours on NVIDIA Hopper during a kernel-optimization task. He claims the model continued finding substantial improvements after its initial exploration phase.

Beta access and open-source questions

Muse Code can be installed with one command and includes a contributor tier, according to Zuckerberg. The announcement does not provide a specific price. Several commenters asked whether access would use a subscription model or API billing, while another questioned whether the contributor tier was available to everyone.

A separate question asked whether Muse Code would be open source. Zuckerberg responded that he would have more to share “soon,” without providing further details.

Muse Spark 1.2 is presented as an intermediate update, with Zuckerberg describing larger models as forthcoming. For now, the beta’s claims rest largely on company-provided benchmarks and testing examples, while the product’s long-running agents and local audit log distinguish its announced workflow from a basic code-completion tool.

Source: X

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community