The Agent Is Only as Reliable as the Intermediate

YuE2's agentic workflow does not add a hidden music-making model behind the released generator. It adds an orchestration layer around the score, request, artifact, and listening interfaces. This chapter explains what that layer can responsibly do and how to diagnose its failures: an unsupported notation construct, a mismatched plan, a truncated result, an upstream transcription error, or an audio judgment that was never actually performed. The practical goal is to inspect an agent-produced change and choose the next check instead of mistaking fluent instructions or a playable file for success.

The principle is simple: an agent can make a symbolic intermediate easier to edit, but it cannot make an invalid intermediate reliable by describing it confidently. YuE2's score-centered design gives the agent a place to state musical intent and a set of artifacts against which to check that intent. The agent remains accountable to those interfaces.

What the skill adds—and what it does not

The packaged yue2-music skill teaches an agent to choose among generation modes, retain an original plan, transcribe and prepare covers, inspect and edit ABC, check musical invariants, decode cached latents, and organize listening comparisons. Those are valuable capabilities because a musical request usually crosses several representations. A person may say “keep the melody, reharmonize the chorus, translate the lyrics, and let the arrangement become jazz.” An agent can turn that request into a bounded score change, a revised request, a comparison record, and a sequence of checks.

But the public workflow does not replace YuE2's generator with a separate agentic-generation checkpoint. The documented agent works between planning and regeneration: it reads the score and the stated constraints, writes a new score or request, passes that result through the released interfaces, and reports what the checks found. The music still comes from YuE2's normal planning, semantic generation, acoustic synthesis, and VAE decoding path.

This division is useful because it assigns ownership. The agent owns the translation of a human request into an explicit proposed intervention and the organization of evidence. The score owns the symbolic musical condition. YuE2 owns the new realization. The comparison tools own the limited invariants they can calculate. Listening owns observations about the sound. When the roles are kept separate, the agent can be helpful without being treated as an authority over every downstream outcome.

A bounded agent brief should therefore require more than a new ABC file. It should identify the intended change, the voices and passages involved, the facts that must remain fixed, the permitted lyric/style changes, and the evidence expected after editing. The skill's guidance calls for changed files, revised prompts or lyrics when needed, an explanation or edit manifest, and invariant results. That record allows another person to inspect not only what the agent produced, but also which interpretation of the request led to it.

A useful manifest might say: preserve every vocal pitch and duration, preserve instrumental melody and section order, leave lyrics and tempo unchanged, change chord symbols only in the chorus, and report the comparison result.