Agent Architectures¶

Design patterns for AI agents and multi-agent systems.

Single Agent Architecture¶

graph TB
    USER[User] -->|Query| AGENT[Agent]
    AGENT -->|Retrieve| RAG[RAG Pipeline]
    AGENT -->|Execute| TOOLS[Tools]
    AGENT -->|Generate| LLM[LLM]
    AGENT -->|Response| USER

Multi-Agent Patterns¶

Pattern: Orchestrator¶

Central coordinator managing specialized agents.

graph TB
    ORCH[Orchestrator] -->|Delegate| A1[Research Agent]
    ORCH -->|Delegate| A2[Code Agent]
    ORCH -->|Delegate| A3[Review Agent]
    A1 -->|Result| ORCH
    A2 -->|Result| ORCH
    A3 -->|Result| ORCH

Pattern: Pipeline¶

Sequential processing through specialized agents.

graph LR
    A1[Parse] -->|Output| A2[Process]
    A2 -->|Output| A3[Validate]
    A3 -->|Output| A4[Format]

Pattern: Debate¶

Multiple agents propose and critique solutions.

RAG Architecture¶

graph TB
    Q[Query] -->|Embed| EMB[Embeddings]
    EMB -->|Search| VS[Vector Store]
    VS -->|Context| LLM[LLM]
    LLM -->|Response| R[Response]

Chunking Strategies¶

Strategy	Use Case	Chunk Size
Fixed	Simple docs	512 tokens
Semantic	Technical docs	Variable
Hierarchical	Long documents	Parent/child

Tool Design¶

Tool Best Practices

Clear, specific descriptions
Minimal required parameters
Idempotent when possible
Graceful error handling

@tool
def search_database(query: str, limit: int = 10) -> list[dict]:
    """Search the database for records matching query.

    Args:
        query: Natural language search query
        limit: Maximum results to return (default: 10)

    Returns:
        List of matching records with id, title, and score
    """
    ...

Anti-Patterns¶

Avoid These

Agents with unclear responsibilities
Missing guardrails/validation
No fallback for tool failures
Unbounded agent loops