claude-agents

Custom Claude Code agents are specialized instances configured with specific tool permissions and pre-written instructions. Define agent metadata (name, description, tools, model, constraints) in frontmatter configuration to enable focused, autonomous task execution with controlled capabilities for workflows like code review, security analysis, or test generation.

Build task-oriented agents by assigning them specialized roles—like `code-reviewer` or `security-analyzer`—with read-only and grep tools to scan for patterns autonomously. Configure iteration limits and model selection (Haiku, Sonnet, Opus) to balance detection depth against cost while executing multi-step analysis workflows.

Yes. Each agent supports model selection across Claude's lineup (Haiku, Sonnet, Opus) and iteration limits tailored to the task. Choose based on capability requirements and cost constraints—Haiku for lightweight scanning, Sonnet for balanced tasks, Opus for complex reasoning.

Define an agent with read-only permissions by specifying tool access in its configuration metadata. Provide task-specific instructions like "read and update project documentation" and include only the tools needed (Read, Glob) to restrict the agent's actions to the defined scope.

Agents execute only with explicitly granted tools; overly restrictive permissions may limit autonomy, while broad permissions reduce safety. Balance capability against control by clearly defining which tools (file operations, grep patterns, globbing) each agent role requires for its specific workflow.

Basic familiarity with Claude's strengths helps, but agent configuration templates standardize setup. Focus on defining the agent's role, required tools, and domain-specific instructions; the template handles the rest of the metadata and constraint structure.