Subagents

Each subagent is essentially a Markdown file whose header contains the name, description, list of available tools and model selection, with the system prompt below.

The Subagents pattern isolates complex subtasks through independent contexts; a subagent's lengthy intermediate process does not affect the main agent's decision space, and the main agent receives only refined conclusions.

Anthropic's multi-agent research system shows that the subagent pattern is markedly effective in tasks requiring multi-step exploration, but token consumption increases significantly, so it should be avoided for simple tasks.

Production-grade agents typically combine five strategies: Prompt Caching, Compaction, Tool-result Clearing, Retrieval, and Subagents.

The recommended scope of claude-code-setup covers five categories: Hooks, Subagents, Skills, MCP Servers, and Plugins.

The context window refers to the maximum number of tokens a large model can process in a single inference, covering the sum of the input prompt and the output completion, and content exceeding the window will be truncated or cause the call to fail.

Subagents are a multi-agent collaboration mechanism; Claude Code can decompose complex tasks and dispatch multiple sub-agents to complete them separately before aggregating the results.

A single Agent producing and validating its own output carries the risk of self-confirmation bias, as the model tends to approve the output it just generated; a Reviewer Agent with an independent context is more likely to find problems.