Meta pairs Muse Spark 1.2 with its own Muse Code agent

Meta shipped Muse Code in beta alongside Muse Spark 1.2. It is a terminal coding agent meant to plan changes across large repositories, write code, and validate results, not merely autocomplete lines in an editor.

Muse Code keeps the session alive with subagents and an event log

The agent runs a simple loop, but keeps async background subagents alive for the whole session. Instead of spawning a helper for every subtask, they gather context continuously and report back only when they have something useful. The runtime appends every model call, tool run, approval, and edit to a local event log, so a crashed session can resume exactly where it stopped.

Bundled skills include /plan for an approval-gated plan, /grill to pressure-test that plan, and /goal to keep work aimed at completion. Muse Spark 1.2 is a coding-focused update over 1.1 with heavier training on generation, debugging, codebase understanding, and end-to-end workflows. Meta co-trained the model with the Muse Code harness: rejection-sampled trajectories, recipe work on goals, compaction, and subagents, plus toolset integration.

Teams no longer buy the model and the harness as separate toys

Willison put the point bluntly: the decisive model trait now is long-sequence agentic tool calling, not a smarter chat persona. Meta underlines that by shipping its own coding agent as part of making long runs work at all.

For engineering teams, the buying unit changes. You are not only picking an API model and a separate harness. You are buying a pair tuned together. A case study shows iterative GPU kernel optimization across more than 1,000 tool calls and runs up to 24 hours on NVIDIA Hopper (KDA and MLA kernels). The agent writes, compiles, profiles, and improves a baseline without simply wrapping a third-party kernel library.

Pricing comes as two model IDs. muse-spark-1.2 is $1.25 per million input tokens and $4.25 per million output. muse-spark-1.2-contributor drops to $0.10 / $0.20 if you let Meta use the data to improve products. That is an order-of-magnitude discount, not a cosmetic promo.

A long-run demo says little about a company review queue

The vendor path is clean: long agent loops, restart-safe logs, co-training. Production still hinges on who approves /plan output, how event-log edits are audited, and whether background subagents can touch secrets outside a sandbox.

Contributor pricing is the second trap. Cheap tokens tempt side projects and corporate PoCs, but consent to product improvement on your prompts and diffs is a governance call, not a line item. Without a data policy, a bargain API becomes a quiet export of codebase context into the vendor training loop.

Merge ownership and contributor traffic will show seriousness

Watch whether teams keep Muse Code off main branches unless a human approves the plan, or whether the agent lands straight in CI. A second signal is how much non-hobby traffic sits on the contributor tier. A third is out-of-demo performance: whether long-horizon runs hold up on foreign monorepos the way the kernel study claims.

Muse Code is in beta on macOS and Linux via install script, and Muse Spark 1.2 is available in Muse Code and the Meta Model API with expanded global access. Claude Code, Codex, and other agentic harnesses already opened this market. Meta is not first, but it is baking model and agent as one product.

Lilith's verdict

Meta is not selling smarter autocomplete. It is selling a model-plus-harness pair baked for long-horizon runs, with a cheap contributor ticket paid in your diffs.