AI 程式開發工具
Meta Launches Muse Code: Persistent Background Agents and Event Logs Support Long-Running Coding Tasks
Muse Code packages Muse Spark 1.2 as a terminal agent capable of working on large codebases in parallel within isolated worktrees. Its append-only event log can reconstruct state after a process interruption, though the performance figures still come primarily from Meta itself.
Meta opened the Muse Code beta on August 5, launching the new Muse Spark 1.2 model alongside a terminal-agent execution framework. Rather than merely generating code, it first explores the codebase, plans changes, invokes tools, runs tests, and continues revising its work based on the results. Large tasks can also be delegated to multiple sub-agents working in parallel in isolated Git worktrees, preventing them from directly touching the user’s current working copy.
The most notable architectural features are persistent background agents and an append-only event log. A background agent continuously retains the context gathered throughout the session instead of rebuilding it for every subtask. Model calls, tool results, approval actions, and file modifications are written sequentially to a local log. If the process crashes or restarts, the executor can restore the work from the event sequence, reducing the cost of rerunning an entire long-running agent task because of a single point of failure.
Meta says Spark 1.2 and Muse Code were co-trained: earlier-generation models created challenging programming environments and helped evaluate candidate solutions, with the results then converted into new training data. In the company’s Terminal-Bench 2.1 evaluation, Spark 1.2 scored 82.9%, below Opus 5 at 86.7% but above several other tested competitors. In another demonstration, the agent made more than 1,000 tool calls on NVIDIA Hopper over runs lasting as long as 24 hours, repeatedly rewriting Triton kernels.
These results should not be treated as directly equivalent to day-to-day maintenance quality. The agent frameworks and reasoning configurations used for the different models were not fully consistent, and GPU-kernel tasks are easier to validate through performance measurements than typical enterprise codebases. Community testing has also uncovered container login issues and shown that simply porting its system prompt can substantially alter cost and completion time. Engineering teams should next examine whether the event log is auditable, how failures by sub-agents are rolled back, and how permissions, secrets management, and spending limits are handled during long-running executions.