Every hour an engineer spends babysitting an AI coding session through context resets is an hour not spent on the work that actually ships.

Repository-scale coding has been the wall most agents can’t climb

Most AI coding tools collapse the moment a task spans more than a few files or requires holding state across dozens of tool calls. Engineers end up breaking complex refactors into manual chunks, restarting sessions, and stitching outputs together themselves.

A terminal agent that stays alive long enough to matter

Meta AI Releases Muse Code (Beta) installs on macOS or Linux with a single curl command, then operates inside your terminal as a persistent coding agent that plans changes, writes code, and validates results across large repositories. Async background agents remain active for the full session rather than spinning up per task. A local append-only event log records every model call, tool run, approval, and edit, making each session replay-exact and restartable if something breaks mid-run.

Infrastructure engineers have the most to gain right now

This tool is built for professionals whose work breaks current agent tools before lunch:

• GPU and HPC engineers who need a single agent to run kernel optimization across thousands of tool calls over hours, not minutes
• Platform engineers managing large-scale migrations who currently split refactors across multiple sessions and lose thread continuity each time
• Senior SaaS developers responsible for long-running bug triage who spend more time restarting tools than reading output

The underlying model, Muse Spark 1.2, was co-trained with the agent harness itself, meaning the model and the scaffolding were built to work together rather than bolted together after the fact.

The 24-hour agent session is a direct shot at Anthropic’s Claude Code

Meta published a case study of Muse Code running over 1,000 tool calls during a single kernel optimization session lasting up to 24 hours, a benchmark that directly challenges the session-length limitations users have reported with Claude Code and similar tools. If the async architecture holds at scale, the gap between what agent tools promise and what they actually deliver in production starts to close in a meaningful way.

What you can do with it today

• Run repository-wide refactors or language migrations without manual session restarts
• Execute long-running bug triage with a persistent agent that retains full context
• Generate test coverage across large codebases in a single supervised session
• Optimize GPU kernels using extended multi-hour agentic runs via the terminal

Access Muse Code through the beta install script or call Muse Spark 1.2 directly through the Meta Model API.

The pricing picture is still forming

Pricing not listed for Muse Code beta — check our directory for updates as the model API tier details become public.

One real constraint worth naming

Meta has not released downloadable weights for Muse Spark 1.2, so every call routes through a hosted dependency you do not control, which matters for teams with strict data residency requirements.

What else is in this space

Claude Code from Anthropic covers similar repository-scale agentic coding through the terminal and has a longer public track record. GitHub Copilot Workspace targets the same planning-to-PR workflow but stays inside the browser rather than the terminal.

Agentic coding is moving from demo to production infrastructure

The shift from single-file assistants to session-persistent coding agents is happening faster than most teams have updated their review and approval workflows to match. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.