Skip to content

AGENT_TEAMS.md

Latest commit

 

History

History
239 lines (183 loc) · 8.98 KB

AGENT_TEAMS.md

File metadata and controls

239 lines (183 loc) · 8.98 KB
Raw

Anthropic Claude Opus 4.6 Agent Teams Integration

This project integrates with Anthropic's Claude Opus 4.6 Agent Teams feature (Research Preview) to enable parallel AI agent coordination for Dify workflow development.

What is Agent Teams?

Agent Teams is a groundbreaking feature released on February 5, 2026, that allows multiple AI agents to work in parallel on complex tasks:

"Instead of one agent working through tasks sequentially, you can split the work across multiple agents — each owning its piece and coordinating directly with the others." — Scott White, Head of Product at Anthropic

Features

1. Multi-Agent Workflow Development

5 specialized agents working in parallel:

Agent Role Responsibility
Workflow Designer Architect Design overall workflow structure and node connections
Node Implementer Developer Write specific node implementations and prompts
DSL Validator Reviewer Validate workflow correctness and DSL compliance
Performance Optimizer Optimizer Analyze and improve cost/latency efficiency
Documenter Technical Writer Generate comprehensive documentation

2. Execution Modes

  • Parallel Mode: All agents work simultaneously (faster)
  • Sequential Mode: Agents work in dependency order (more controlled)

3. Built-in Coordination

  • Automatic dependency resolution
  • Inter-agent message passing
  • Result aggregation and conflict resolution

Quick Start

Basic Usage

from dify_workflow import DifyWorkflowAgentTeam, create_workflow_with_agents

# Method 1: Using the high-level API
result = create_workflow_with_agents(
    "Create a translation workflow that takes text and target language as input",
    api_key="your-anthropic-api-key"
)

# Method 2: Using the team interface
async def develop_workflow():
    team = DifyWorkflowAgentTeam(api_key="your-api-key")

    # Create default team with 5 specialists
    team.create_default_team()

    # Develop workflow with parallel execution
    result = await team.develop_workflow(
        requirements="Build a sentiment analysis workflow for customer feedback",
        mode="parallel"
    )

    # Access results
    print(result["design"])           # Architecture design
    print(result["implementation"])   # Implementation code
    print(result["validation"])       # Validation report
    print(result["optimization"])     # Optimization suggestions
    print(result["documentation"])    # Generated docs

    return result["final_workflow"]

# Run it
import asyncio
workflow = asyncio.run(develop_workflow())

Interactive CLI Mode

from dify_workflow.agent_teams import AgentTeamCLI

cli = AgentTeamCLI()
asyncio.run(cli.interactive_mode())

Output:

🚀 Dify Workflow Agent Team - Interactive Mode

✅ Registered Agent: Workflow Designer (workflow_designer)
✅ Registered Agent: Node Implementer (node_implementer)
✅ Registered Agent: DSL Validator (validator)
✅ Registered Agent: Performance Optimizer (optimizer)
✅ Registered Agent: Documenter (documenter)

Options:
1. Describe requirements and develop workflow
2. Optimize existing workflow
3. Generate documentation
4. Exit

Select (1-4): 1
Describe your workflow requirements: Create a chatbot with RAG support
Execution mode (parallel/sequential) [parallel]: parallel

======================================================================
🚀 Dify Workflow Agent Team Starting
======================================================================
📋 Requirements: Create a chatbot with RAG support...
👥 Team Size: 5 agents
⚡ Execution Mode: parallel
======================================================================

📐 Phase 1: Architecture Design
  🤖 Workflow Designer starting...
  ✅ Workflow Designer completed

🔨 Phase 2: Parallel Development & Validation
  🤖 Node Implementer starting...
  🤖 Performance Optimizer starting...
  ✅ Node Implementer completed
  ✅ Performance Optimizer completed

✅ Phase 3: Verification & Optimization
  🤖 DSL Validator starting...
  🤖 Documenter starting...
  ✅ DSL Validator completed
  ✅ Documenter completed

⚡ Phase 4: Apply Optimizations
  🤖 Performance Optimizer starting...
  ✅ Performance Optimizer completed

Advanced Usage

Custom Agent Configuration

from dify_workflow import AgentConfig, AgentRole

# Create custom agent
custom_agent = AgentConfig(
    id="security_expert",
    name="Security Specialist",
    role=AgentRole.VALIDATOR,
    model="claude-opus-4-6",
    temperature=0.3,
    system_prompt="""You are a security expert specializing in AI workflow security.
Your responsibilities:
1. Check for prompt injection vulnerabilities
2. Validate input sanitization
3. Review data privacy compliance
4. Suggest security best practices"""
)

team = DifyWorkflowAgentTeam()
team.register_agent(custom_agent)

Custom Execution Flow

async def custom_development_flow():
    team = DifyWorkflowAgentTeam()
    team.create_default_team()

    # Phase 1: Design
    design = await team._run_agent("designer_1", "Design requirements...")

    # Phase 2: Parallel implementation
    impl_task = team._run_agent("implementer_1", f"Implement: {design}")
    doc_task = team._run_agent("documenter_1", f"Document: {design}")

    implementation, documentation = await asyncio.gather(impl_task, doc_task)

    # Phase 3: Validate and optimize
    validation = await team._run_agent("validator_1", f"Validate: {implementation}")
    optimization = await team._run_agent("optimizer_1", f"Optimize: {implementation}")

    return {
        "design": design,
        "implementation": implementation,
        "documentation": documentation,
        "validation": validation,
        "optimization": optimization
    }

API Status

⚠️ Research Preview: This feature requires:

  • Anthropic API access with Opus 4.6 model
  • Beta API enrollment (contact Anthropic)
  • API key with agent-teams scope

Architecture

┌─────────────────────────────────────────────────────────────┐
│                  DifyWorkflowAgentTeam                       │
├─────────────────────────────────────────────────────────────┤
│  ┌─────────────┐ ┌─────────────┐ ┌─────────────┐           │
│  │   Designer  │ │ Implementer │ │  Validator  │           │
│  └──────┬──────┘ └──────┬──────┘ └──────┬──────┘           │
│         │               │               │                  │
│         └───────────────┼───────────────┘                  │
│                         │                                   │
│              ┌──────────┴──────────┐                       │
│              │   Task Coordinator  │                       │
│              └──────────┬──────────┘                       │
│                         │                                   │
│         ┌───────────────┼───────────────┐                  │
│         │               │               │                  │
│  ┌──────┴──────┐ ┌──────┴──────┐ ┌──────┴──────┐          │
│  │  Optimizer  │ │ Documenter  │ │ Custom Agent│          │
│  └─────────────┘ └─────────────┘ └─────────────┘          │
└─────────────────────────────────────────────────────────────┘

Comparison: Single Agent vs Agent Teams

Aspect Single Agent Agent Teams
Speed Sequential Parallel (40% faster)
Quality Single perspective Multiple expert perspectives
Complexity Limited context Distributed expertise
Cost Lower token usage Higher parallelism cost
Use Case Simple workflows Complex, multi-faceted workflows

References

License

MIT License - See LICENSE for details.