Scale Details Muse Code: Co-training Model and Agent for Better Tool Use

alexandr_wang · x · 2026-08-06

Scale AI elaborated on the core technical and architectural features of Muse Code:

Co-training: By co-training the model and the agent together, it achieves higher first-attempt success on tool calls, cleaner plan execution, and significantly less re-prompting.
Long Context: Supports a massive 1 million token context window.
Parallel Architecture: The harness architecture allows multiple specialized agents to coordinate in parallel on the same task. They persist across sessions to build context, and for complex work, sub-agents fan out into isolated worktrees to keep the working copy clean.
Real-world Performance: In a test optimizing kernels on NVIDIA Hopper, the agent ran for 24 hours making over 1,000 tool calls, continuously finding substantial improvements well beyond the initial exploration phase.

More from coding & agent