Claude Agents

Claude Code agents specialize work that would otherwise fill the main conversation with search results, codebase notes, or review findings. A good agent definition states when to use it, the tools it may access, and the output it should return. The entries here focus on research, review, and development workflows with links to their actual source repositories.

Start with a bounded request such as finding the implementation of an API or reviewing a specific diff. State the files or directories in scope and the format of the result. A short report with paths, findings, and open questions is easier to act on than an unrestricted exploration of the entire repository.

When comparing definitions, look for explicit responsibilities and tool access. A research worker that only reads files has different needs from an implementation worker. Check whether the source supplies a single definition or an orchestration package, because the installation and the way you invoke it can differ.

In Claude Code documentation, a custom subagent is a specialized agent that works in its own context window. Community projects often use “agent” for an entire plugin or orchestration workflow and “subagent” for an individual worker definition. This directory keeps both terms so users can browse by intent; the subagents page goes deeper into individual Markdown templates and placement.

The labels on community repositories are not a separate permission system. Read the actual Markdown definition and plugin configuration to understand what runs. A package advertised as an agent framework may install several custom subagents, commands, and skills together, while a single review agent may be only one file.

Use the agents collection to compare complete workflows, then use the subagent templates when you want to tune one role. Keep the handoff explicit: what the worker receives, what it returns, and who acts on its findings. This also makes it easier to see when extra delegation adds work without improving the answer.

Plugin-based agents can be installed with the marketplace commands on each card. For a single custom agent, save a Markdown file with name, description, tool access, and instructions in `.claude/agents/` for one project or `~/.claude/agents/` for your user account. Review permissions before placing it there. The downloadable example below starts read-only so you can adapt it safely.

The example below restricts a reviewer to Read, Grep, and Glob. Save the file and confirm its name and description in the frontmatter before trying it; /agents provides setup guidance. Give it a small set of changed files and check that it reports issues with evidence rather than editing the project.

Keep a project-specific definition in version control so changes to the role can be reviewed. Add dependencies only when a real task requires them. If the agent needs command execution later, document the expected commands and recheck its behavior instead of granting a broad tool list just to make setup easier.