Claude Code Agent Patterns

Claude Code Agent Patterns

Guide for designing and implementing Claude Code agents with proven architectural patterns, optimal tool selection, and effective extended context integration.

When to Use This Skill

Activate when:

• Creating new Claude Code agents
• Designing agent architecture and responsibilities
• Selecting appropriate tools for agents
• Integrating extended context with agents
• Implementing multi-agent coordination
• Troubleshooting agent behavior

Core Architecture Patterns

1. Single-Purpose Specialist

Use Case: Agents focused on one specific domain or capability.

---
name: security-scanner
description: "Specialized agent for identifying security vulnerabilities"
tools: [Read, Glob, Grep, Write]
model: sonnet
---

Characteristics:

• Clear, focused responsibility
• Deep domain expertise
• Optimized tool set
• Rich extended context

Best For: Analysis tasks, specialized assessments, domain-specific operations

2. Workflow Orchestrator

Use Case: Agents coordinating multiple tasks or other agents.

---
name: ci-cd-orchestrator
description: "Coordinates complete CI/CD pipeline execution"
tools: [Task, Bash, Read, Write, TodoWrite]
model: opus
---

Characteristics:

• Manages complex multi-step processes
• Coordinates other agents via Task tool
• Handles error recovery and rollback
• Maintains process state

Best For: Complex workflows, multi-agent coordination, pipeline orchestration

3. Context-Aware Adapter

Use Case: Agents that adapt behavior based on project context.

---
name: language-aware-formatter
description: "Code formatter adapting to project language and style"
tools: [Read, Glob, Bash, Write, Grep]
model: sonnet
---

Characteristics:

• Detects project characteristics automatically
• Adapts behavior based on context
• Supports multiple scenarios
• Maintains consistent interface

Best For: Multi-language tools, adaptive analysis, context-sensitive operations

4. Resource-Optimized Processor

Use Case: Agents handling large datasets or resource-intensive operations.

---
name: large-codebase-analyzer
description: "Efficiently analyzes large codebases with incremental processing"
tools: [Read, Glob, Bash, Write, TodoWrite]
model: sonnet
---

Characteristics:

• Implements incremental processing
• Manages memory and resources
• Provides progress tracking
• Supports resumable operations

Best For: Large codebase analysis, batch processing, long-running operations

Model Selection Guidelines

Haiku

• Simple validation tasks
• File operations
• Basic path checking
• Quick utility operations

Sonnet

• Complex reasoning and analysis
• Domain-specific expertise
• Code generation and modification
• Most specialist agents

Opus

• Multi-agent coordination
• Workflow orchestration
• Complex decision-making
• High-stakes operations

Tool Assignment Patterns

Analysis-Focused

tools: [Read, Glob, Grep, Write]

Information gathering and reporting

Orchestration-Focused

tools: [Task, Bash, TodoWrite]

Process coordination and execution

Development-Focused

tools: [Read, Edit, MultiEdit, Bash, Write]

Code modification and testing

Web-Focused

tools: [WebFetch, WebSearch, Read, Write]

External information retrieval

Extended Context Architecture

Hierarchical Knowledge Structure

extended_context/
├── domain-knowledge/
│   ├── core-concepts/
│   ├── best-practices/
│   └── common-patterns/
├── troubleshooting/
│   ├── common-issues/
│   └── resolution-guides/
└── reference-materials/
    └── tool-specifications/

Cross-Referenced Context

# In security-patterns.md
Related contexts:
- ../coding-standards/security-practices.md
- ../vulnerability-databases/known-issues.md

Layered Context

1. Core Layer: Fundamental concepts
2. Domain Layer: Specific knowledge areas
3. Application Layer: Implementation guides
4. Troubleshooting Layer: Problem resolution

Agent Communication Patterns

Message Passing

Context passed to next agent:
- operation-id: unique identifier
- previous-results: structured output
- user-context: original request

Shared Context

Shared context location: extended_context/shared/
- project-state.md
- user-preferences.md
- session-context.md

Quality Assurance Patterns

Self-Validation

Validation steps:
1. Check output format compliance
2. Verify required information present
3. Validate against known constraints
4. Perform consistency checks

Peer Review

Review workflow:
1. Primary agent completes analysis
2. Review agent validates quality
3. Synthesis agent combines perspectives
4. Final output with confidence scores

Anti-Patterns to Avoid

Monolithic Agent

Problem: Single agent handling too many responsibilities Solution: Split into focused specialist agents

Tool Overload

Problem: Agent with too many tools Solution: Assign minimal needed tool set

Context Redundancy

Problem: Duplicate information across agents Solution: Create shared context areas

Coordination Chaos

Problem: Agents without clear communication Solution: Implement structured coordination patterns

Agent Template

---
name: agent-name
description: "Clear, concise purpose statement"
model: sonnet  # or haiku/opus based on complexity
tools: [Read, Write, ...]
---

# Agent Name

## Purpose
[What this agent does and when to use it]

## Capabilities
- [Capability 1]
- [Capability 2]

## Extended Context Integration
References:
- extended_context/[category]/[file].md

## Input Requirements
- [Required input 1]
- [Required input 2]

## Output Format
[Expected output structure]

## Anti-Fabrication Requirements
- Base all outputs on actual analysis
- Execute validation tools before claims
- Mark uncertain information appropriately

Key Principles

• Single Responsibility: One clear purpose per agent
• Minimal Tools: Only tools needed for the task
• Rich Context: Comprehensive extended context
• Clear Communication: Structured coordination patterns
• Self-Validation: Verify outputs before completion
• Consistent Naming: lowercase-hyphen format