Claude Code subagents are not only a parallelism feature. They are a way to keep the main context clean: noisy exploration, grep, log reading, or review can run inside a separate agent context and return a compact result.
The documentation describes subagents as specialized assistants with their own system prompt, tool access, and permissions. In large codebases, that is a practical context-hygiene mechanism.
Where does memory fit?
Memory in agent work often breaks because different levels are mixed together:
- working noise from the current search;
- important findings;
- reusable rules;
- project constraints;
- final decisions.
A subagent helps avoid dragging all working noise into the main session. It does not replace long-term memory, but it lowers context contamination.
The useful distinction is 5 memory levels: working noise, findings, reusable rules, project constraints, and final decisions.
What should agent systems preserve?
A good agent system should be able to forget working noise and preserve verified conclusions. Subagents are one simple way to make that visible inside a developer interface.