persistent sub-agents

Persistent sub-agents are a way for a program handling one large, complex task to break that task into pieces and assign each piece to a dedicated helper agent that sees it through to completion.

Think of a general contractor on a big construction site who assigns plumbing, electrical, and finishing work to separate crews. These crews aren't day laborers who show up once and disappear—they're full-time staff who stay on their section of the site, tracking progress until the job is done. In the same way, a coding program handling changes across many files doesn't just answer one question and stop. Instead, a dedicated sub-agent for each part keeps planning, fixing, and verifying all the way through.

This structure matters because tasks like software development unfold over many steps to reach completion, and having to re-explain the full context every single time would be wasteful. By dividing the work while letting each sub-agent retain its own context, changes can be made consistently even across a large codebase.

Meta introduced its coding agent Muse Code, describing it as using "persistent sub-agents to divide up and handle tasks." A common misunderstanding is treating sub-agents as just an extra Q&A channel, but the key word is "persistent": rather than answering once and being done, they stay engaged, holding onto context throughout the planning, implementation, and verification process across a large repository.