Skip to main content

Anthropic Agent Patterns & Multi-Agent Architectures

Research Date: December 2025 Purpose: Document agent architecture patterns and coordination strategies Sources: Anthropic Agent Skills research, engineering blogs, CODITECT implementation

Executive Summary​

Agent Skills architecture enables scalable, modular agent systems through progressive disclosure, metadata-driven discovery, and hierarchical coordination patterns.

1. Agent Skills Architecture​

Core Principles​

  1. Progressive Disclosure - Load content on-demand
  2. Metadata First - Lightweight discovery
  3. Modular Design - Composable agent capabilities
  4. Hierarchical Organization - Nested skill trees

Three-Tier Architecture​

  1. Metadata Layer (Always Loaded)

    • Skill name and description
    • Keywords and categories
    • Prerequisites and dependencies

  2. Full Content Layer (Conditionally Loaded)

    • Detailed instructions
    • Implementation patterns
    • Examples and templates

  3. Referenced Resources (As-Needed)

    • Documentation files
    • Code examples
    • External references

Benefits​

  • Minimal baseline token consumption
  • Unbounded skill catalog potential
  • Fast skill discovery
  • Focused context loading

2. Multi-Agent Coordination​

Coordination Patterns​

1. Hierarchical (Lead + Workers)​

  • Lead Agent: High-level planning and coordination
  • Worker Agents: Specialized task execution
  • Communication: Via structured messages
  • State: Shared or partitioned

2. Collaborative (Peer-to-Peer)​

  • Equal Agents: Shared responsibility
  • Negotiation: Task allocation through discussion
  • State: Shared knowledge base
  • Consensus: Agreement on decisions

3. Sequential (Pipeline)​

  • Ordered Agents: Each handles specific phase
  • Handoff: Output → Input chain
  • State: Accumulated through pipeline
  • Validation: Each stage validates input

Sub-Agent Patterns​

From Anthropic's "Effective Harnesses for Long-Running Agents":

When to Use Sub-Agents​

  • Complex problems requiring specialized expertise
  • Context window approaching limits
  • Early in a conversation or task
  • Parallel workstreams

Sub-Agent Benefits​

  • Context Isolation: Fresh context per sub-agent
  • Specialization: Focused expertise
  • Scalability: Unlimited parallel agents
  • Efficiency: Condensed summaries returned

Implementation Pattern​

## Sub-Agent Invocation

1. Lead agent identifies specialized task
2. Spawns sub-agent with specific system prompt
3. Sub-agent executes with clean context
4. Returns condensed summary (1,000-2,000 tokens)
5. Lead agent integrates summary into plan

3. Agent Communication​

Message Passing​

  • Structured Format: JSON or Markdown
  • Clear Semantics: Action, data, metadata
  • Validation: Schema-based verification
  • Logging: All inter-agent messages

Shared State​

  • Centralized: Single source of truth
  • Partitioned: Agent-specific state
  • Synchronized: Eventual consistency
  • Versioned: Conflict detection

Coordination Mechanisms​

  • Task Queue: Pending work items
  • Event Bus: Pub/sub notifications
  • Locks: Prevent race conditions
  • Checkpoints: Recoverable state

4. CODITECT Agent Framework​

Agent Activation Pattern​

# Agent discovery
agents = search_agents(task_description)

# Activation check
if not is_activated(agent_name):
    request_activation(agent_name, reason)

# Agent invocation (Task Tool Proxy Pattern)
result = Task(
    subagent_type="general-purpose",
    prompt=f"Use {agent_name} subagent to {task}"
)

Agent Specialization​

  • Project Agents: Project management, planning
  • Research Agents: Analysis, documentation
  • Development Agents: Code generation, refactoring
  • QA Agents: Testing, validation, review
  • Security Agents: Audits, compliance, scanning
  • Deployment Agents: CI/CD, infrastructure

Coordination Strategy​

  1. Orchestrator: High-level planning
  2. Specialist Agents: Domain expertise
  3. Integration: Results consolidation
  4. Validation: Quality gates

5. Best Practices​

Agent Design​

  • Single responsibility per agent
  • Clear input/output contracts
  • Well-defined expertise boundaries
  • Composable with other agents

Coordination​

  • Minimize inter-agent messages
  • Use structured message formats
  • Validate all agent interactions
  • Log coordination events

Performance​

  • Lazy load agent capabilities
  • Cache agent metadata
  • Parallel execution where possible
  • Monitor agent utilization

Error Handling​

  • Graceful degradation
  • Retry with exponential backoff
  • Circuit breaker pattern
  • Fallback agents

6. Source References​

  1. Equipping Agents with Agent Skills
  2. Effective Harnesses for Long-Running Agents
  3. Effective Context Engineering for AI Agents
  4. CODITECT agent framework implementation

Last Updated: December 2025 Status: Production implementation with 50+ specialized agents