What “persistent background agents” means

Meta describes Muse Code as a simple main-agent loop supported by async background agents. These agents remain active through the session rather than being created only for a single subtask. They can take next steps independently and choose when to report back to the main agent.

Persistent background workConceptual · not to scale
Three workers are shown to explain fan-out, not to claim a fixed product limit.

Persistence can reduce duplicate discovery and waiting on difficult tasks, but it also means more work may happen outside the main conversational thread. Review and validation remain essential.

Pattern
Main + background
Official runtime description
Lifecycle
Session-persistent
Not per-small-task only
Execution
Async + parallel
Can advance independently
Fixed count
Not published
Overview page checked

Good evaluation tasks

Task shapePotential parallel workReview focus
Cross-cutting featureExplore UI, data, and test surfacesInterface consistency
Large failing suiteInvestigate independent failure clustersShared root causes
Repository migrationMap packages and compatibility risksIntegration order
Research-heavy fixTrace call paths and documentationEvidence quality

Start with work that has separable questions and a crisp integration test. A vague request does not become well-scoped merely because more agents can work on it.

Questions the overview does not settle

The current product and launch pages do not specify a universal scheduler, token allocator, concurrency ceiling, or a single isolation strategy for every task. Meta’s broader cookbook contains multi-agent patterns, but those examples should not automatically be treated as the hidden implementation of Muse Code itself.

For the model relationship, read Muse Spark 1.2. For the audit and recovery layer that records actions around agent work, continue to event log and resume.