9. Agent Orchestration

Sequential Workflows

Chain agents in sequence where each agent's output becomes the next agent's input. Useful for document processing pipelines.

# Sequential workflow pseudocode
# Agent 1: Extract document metadata
# Agent 2: Classify document type
# Agent 3: Generate summary
# Agent 4: Store results

async def process_document(document):
    metadata = await extract_agent.run(f"Extract metadata from: {document}")
    doc_type = await classify_agent.run(f"Classify: {metadata}")
    summary = await summarize_agent.run(f"Summarize: {document}")
    await storage_agent.run(f"Store: {metadata} | {doc_type} | {summary}")
    return summary

Human-in-the-Loop (HITL)

For critical decisions, agents can pause and request human approval before proceeding.

# HITL pattern
async def approve_expense(amount, reason):
    if amount > 1000:
        # Pause and ask human
        approval = await ask_human(
            f"Approve ${amount} for {reason}?",
            timeout=3600  # Wait up to 1 hour
        )
        if not approval:
            return "Expense rejected by manager"
    return "Expense approved"

Agent Chat Systems

Build multi-turn chat systems where different agents handle different parts of the conversation:

• Triage Agent — Routes to the right specialist
• Support Agent — Answers technical questions
• Billing Agent — Handles invoices and payments
• Escalation Agent — For complex issues requiring human intervention

Orchestration Best Practices

• Define clear handoff criteria between agents
• Pass conversation context explicitly during handoffs
• Set timeout limits for human-in-the-loop steps
• Log all orchestration decisions for audit