AI Agents

Multi-Agent Workflows

Multi-agent workflows let you decompose complex tasks into specialised sub-agents coordinated by an orchestrator.

When to Use Multi-Agent

• The task is too complex for a single prompt
• Different sub-tasks require different models (cost vs. quality tradeoff)
• Sub-tasks can run in parallel
• You need specialised agents with distinct system prompts

Pattern 1: Orchestrator + Workers

[Input]
    ↓
[Orchestrator Agent]
  ↙       ↓       ↘
[Worker A] [Worker B] [Worker C]
  ↘       ↓       ↙
    [Merge Results]
        ↓
   [Final Output]

The orchestrator breaks down the task, delegates to workers, and synthesises results.

Pattern 2: Pipeline

[Research Agent] → [Analysis Agent] → [Writer Agent] → [Review Agent]

Each agent handles one stage, passing enriched data to the next.

Pattern 3: Supervisor with Retry

[Task] → [Worker Agent] → [Supervisor Agent]
                  ↑_____________________________|
                     (re-delegate if quality fails)

Setting Up a Multi-Agent Workflow

1. Add multiple AI Agent nodes to the canvas
2. Connect them with edges (sequential) or via a Merge node (parallel)
3. Use the Set node to package one agent's output as the next agent's input
4. Add a Switch node if the orchestrator needs to conditionally route

Cost Considerations

• Use powerful models (GPT-4o, Claude 3.5) for orchestration
• Use fast, cheap models (GPT-4o-mini, Haiku) for workers
• Monitor token usage in Analytics → AI Costs