Meta wants a seat in your terminal. This week it shipped Muse Code, a beta command-line agent that takes a task, plans the changes, writes the code and then checks its own work. It's Meta's first real shot at Anthropic's Claude Code and OpenAI's Codex — and an admission that the agent race, not just the model race, is where developer mindshare gets won.
So what is it?
Not autocomplete. You point Muse Code at a repository — including a big, messy one — describe what you want, and let it work. It runs on macOS and Linux for now, and installation is one curl command from Meta's dev site.
The interesting part is how it's built. Muse Code keeps a pool of async background agents alive for your entire session rather than spawning a fresh one per task. Jobs get split across parallel sub-agents, and each works inside its own isolated copy of the repo, so two experiments can't stomp on each other's edits.
The flight recorder
Meta's standout design choice, per The Register's coverage: everything gets written to a local, append-only event log — every model call, every tool run, every approval, every edit. Meta calls it replay-exact and restart-safe. Crash mid-task? Resume from the log. Want to know what the agent actually did at 2 a.m.? Replay it. If you're going to hand an AI write access to your codebase, a tamper-evident receipt of its every move is a decent trust offering.
The model underneath
The agent runs on Muse Spark 1.2, a coding-tuned update to 1.1 that Meta says is better at generation, gnarly debugging and understanding whole codebases. Model and tool were trained as a pair — the training data included rejection-sampled runs from the Muse Code harness itself, tuned for how the agent manages goals, compacts context and delegates to sub-agents.
Meta's benchmark flex: 59% on DeepSWE 1.1, which it says beats Grok Build 4.5 and Gemini 3.6 Flash. Vendor numbers always deserve a raised eyebrow, and independent testing will land soon enough. Still, co-training a model with its harness is exactly how the current leaders got good, so the approach itself is credible.
Why this matters
The terminal turned into AI's most contested real estate almost overnight. Coding agents are the one category of AI agent that unambiguously makes money, and Meta — for all the billions it pours into AI — had nothing on that shelf. Now it does, and it's competing on architecture (persistence, isolation, replayability) instead of just vibes.
If you write software, a serious third player is good news regardless of whose logo you prefer. Competition here tends to show up as lower prices and faster fixes.
The catch to watch
Two open questions decide Muse Code's fate. Does it hold up in ugly real-world repos, where every agent looks worse than its benchmarks? And will developers trust Meta with their code — because questions about where usage data goes and what trains the next model aren't going away. The beta's free. The people who live in terminals will render the verdict quickly; they always do.
Image: Jakub Zerdzicki, via Pexels
