Agents

Subagents for isolated, complex workflows

Agents (~/.claude/agents/*.md) are specialized subagents that run in their own context window — separate from the main conversation. They were the original approach to complex multi-step tasks before skills matured.

Current Status

In practice, agents have been largely superseded by skills and direct work in the main conversation. Skills load into your existing context (preserving all accumulated understanding), while agents start fresh in isolation. The main use case for agents now is when you genuinely need isolation — long-running tasks that might pollute the main conversation, or parallel work streams.

Available Agents

analysis-agent
Deep analysis of complex codebase issues requiring multi-file investigation. For tracing data flow, understanding intricate interactions, "why is this happening" questions.
coding-agent
Execute implementation plans by writing code. Follows established patterns, builds functionality step-by-step from plans.
planning-agent
Analyze requirements and create implementation plans with file structure trees, execution tasks, and issue tracking.
research-agent
Search web for implementation examples and patterns. For "how do others implement X" or finding reference implementations.
reference-agent
Analyze library source code in reference/ directory. For understanding how frameworks actually work internally.
testing-agent
E2E testing workflow with Playwright. Test execution, log tailing, network monitoring, result analysis.

Agents vs. Skills vs. Commands

Dimension Commands Skills Agents
Invocation Explicit (/command) Automatic (keyword match) Explicit or automatic
Context Same conversation Same conversation Separate context window
State Shares all context Shares all context Starts fresh
Best for Specific workflows Domain capabilities Isolated heavy work
Location ~/.claude/commands/ ~/.claude/skills/*/ ~/.claude/agents/
The Simpler Abstraction Wins

The evolution from agents to skills to direct work mirrors the broader pattern: a simpler abstraction beats a heavier one. Agents are powerful but heavyweight. Skills are lighter, composable, and share context. Often the best approach is no abstraction at all — just give Claude the context directly and let it work.