August 8, 2026

6 min read

Artificial Intelligence

Meta Muse Code: Terminal Coding Agent Debuts

> Meta shipped Muse Code and Muse Spark 1.2 on August 5, 2026 — a co-trained terminal coding agent with persistent background agents and replay-safe sessions.

ShareX LinkedIn

🎧 Listen — ~6 min

Ready · Meta Muse Code: Terminal Coding

0:00 / 6:00

Verified by Essa Mamdani

Meta entered the terminal coding agent race on August 5, 2026, shipping Muse Code (beta) alongside Muse Spark 1.2, a coding-focused update to its month-old Muse Spark 1.1 model. It's Meta's third named model release in four months and its first purpose-built coding agent, putting Meta Superintelligence Labs in direct competition with Claude Code, Codex CLI, and Gemini CLI.

Here's what actually shipped, how it's built, and whether it's worth adding to your toolchain.

Installing it

Muse Code runs on macOS and Linux via a one-line installer:

curl -fsSL https://dev.meta.ai/install.sh | bash

No IDE plugin, no browser extension — it's a terminal-native agent in the same mold as Claude Code and Codex CLI, powered by Muse Spark 1.2 through the Meta Model API.

Async background agents, not just a single loop

The architectural bet Meta is making is on persistent background agents rather than the more common pattern of spawning a fresh subagent per task. Muse Code runs a simple main agent loop, but background agents stay active for the whole session instead of being created and torn down for each step. They keep working on their assigned piece and decide for themselves when to report back to the main agent.

The practical effect: less redundant context-gathering. If a background agent already scanned a module for a prior step, it doesn't re-scan it for the next one — it holds that state. Meta says this reduces both latency and the amount of manual steering needed on multi-step tasks, which tracks with where the whole agentic coding field has been pushing: the bottleneck isn't raw model capability anymore, it's how well the harness manages state across a long task.

Replay-exact, restart-safe by design

Muse Code keeps a local event log that appends every model call, tool run, human approval, and file edit as it happens. That log is the single source of truth for the session — which means if the process crashes mid-task, Muse Code can resume exactly where it left off rather than restarting the whole plan or losing context.

For anyone who's had a long-running agent session die at hour three of an overnight refactor, this is the feature that matters more than any benchmark number. Long-horizon tasks are only useful if failure doesn't mean starting over.

Built-in skills: /plan, /grill, /goal

Muse Code ships three default skills baked into the harness:

/plan — turns a task description into an approval-gated plan before any code gets written, so you review the approach before the agent starts touching files.
/grill — stress-tests that plan, poking at edge cases and assumptions until it either holds up or gets revised.
/goal — works iteratively toward a stated objective rather than a fixed step list, useful for open-ended tasks where the exact path isn't known upfront.

This is a more structured take on agent guardrails than "just prompt it and hope" — plan-then-execute with an adversarial review step in between is a sensible default for anyone nervous about giving an agent broad repo access.

Muse Spark 1.2: co-trained with the harness, not bolted on

The model and the agent were trained together, not paired after the fact. Meta says training included rejection-sampled harness trajectories — meaning the model learned from actual Muse Code sessions, including how it uses goals, context compaction, and subagents — plus direct integration of the Muse Code toolset during training. That co-training is why Meta claims the model performs best specifically inside Muse Code, versus being a general-purpose model that happens to also do coding.

Three specific training investments stand out:

Long-horizon coding. Muse Spark 1.2 was trained extensively on whole-repository generation, large end-to-end projects, and auto-research tasks — the kind of work that spans dozens of files and hours of wall-clock time, not single-function completions.

Self-improvement loop. Meta used the prior model, Muse Spark 1.1, to generate hard coding environments and instruction-following templates, then had it grade candidate solutions against those requirements. That produced a scalable synthetic training set that Meta says specifically improved 1.2's precision on complex, multi-constraint instructions.

Scaled coding compute. Meta significantly increased training compute allocated to coding tasks and broadened the diversity of training environments, while claiming the model held onto its general-agent capability rather than trading one for the other.

The 24-hour kernel optimization test

The most eye-catching benchmark in Meta's announcement isn't a leaderboard score — it's a live stress test. Meta had Muse Spark 1.2, running inside Muse Code, iteratively optimize GPU kernels for NVIDIA Hopper hardware across more than 1,000 tool calls over up to 24 hours straight, targeting KDA and MLA kernel implementations.

The constraint that makes this meaningful: models were barred from importing third-party kernel libraries like FLA. They had to actually apply kernel-optimization knowledge to hand-write Triton implementations rather than wrap an existing library and call it done. Muse Spark 1.2 reportedly combined standard fusion and tiling techniques with kernel-specific tricks — re-centering the gated cumulative decay at the chunk midpoint for KDA, and reusing the shared KV latent as both K and V for MLA — producing a two-kernel Triton pipeline that beat the PyTorch reference baseline.

That's a genuinely hard, narrow, verifiable task, and a smarter demo than another round of SWE-bench percentage points. It's the kind of long-horizon, tool-heavy, self-correcting work that's supposed to be the entire point of agentic coding tools.

Where this sits in the field

Muse Spark 1.2 arrives on the heels of a packed week: Alibaba's Qwen3.8-Max, DeepSeek-V4-Flash-0731, and GPT-5.6 Sol updates all landed within days of each other, and Meta's own release cadence — Muse Spark in April, Muse Spark 1.1 in July, Muse Spark 1.2 in August — shows a company iterating fast to catch up rather than leapfrog. Independent trackers have placed Muse Spark 1.2 outside the very top tier on standardized Terminal-Bench leaderboards even as Meta's own reported numbers look stronger, which is worth keeping in mind: this is Meta's own benchmark report, not yet independently verified on public leaderboards.

What's not in dispute is the strategic move. Meta didn't just ship a better model — it shipped a coding agent to go with it, on the same day, co-trained end to end. That's the same playbook Anthropic and OpenAI have run with Claude Code and Codex, and it signals Meta sees the harness itself, not just the underlying model, as a competitive surface worth owning.

Getting started

Muse Spark 1.2 is available today through Muse Code and the Meta Model API, with what Meta describes as expanded global access compared to the 1.1 preview. If you're already running Claude Code or Codex CLI daily, Muse Code is worth a trial run specifically on a long, messy, multi-file task — that's where the async background agents and replay-safe event log are designed to earn their keep, not on a quick single-file edit.