Skill מערכות Multi-Agent ל-Claude

מתי להשתמש

"Multi-agent", "Agent team", "Specialized agents", "Agent orchestration", "Sub-agents".

הוראות עבודה

1. Why Multi-Agent

Single agent struggles with complex multi-domain tasks. Multiple specialized agents = better results.

Single Agent

• Generic, knows little about everything.
• Long context = quality drops.

Multi-Agent

• Each specialized.
• Orchestrator delegates.
• Better quality per task.

2. Patterns

א. Orchestrator + Specialists

Orchestrator
├── Researcher Agent
├── Writer Agent
├── Reviewer Agent
└── Publisher Agent

ב. Sequential Pipeline

Agent 1 → Agent 2 → Agent 3 → Output

ג. Debate / Consensus

Agent A (proposes)
Agent B (critiques)
Agent C (judges)

ד. Hierarchical

Manager Agent
├── Team Lead 1
│   ├── Worker 1
│   └── Worker 2
└── Team Lead 2
    ├── Worker 3
    └── Worker 4

3. Frameworks

CrewAI

• Strengths: Easy role-based agents.
• Use: Marketing crew (researcher, writer, editor).

AutoGen (Microsoft)

• Strengths: Multi-agent conversations.
• Use: Coding teams (planner, coder, reviewer).

LangGraph

• Strengths: Graph-based, complex flows.
• Use: Production multi-agent.

Anthropic Sub-agents

• Strengths: Built into Claude Agents SDK.
• Use: Production with Claude.

4. When to Use Multi-Agent

✅ Best For

• Complex projects (research → write → review).
• Multi-domain expertise (legal + finance + marketing).
• Quality through redundancy (multiple reviewers).
• Scalability (parallel agents).

❌ Skip For

• Simple tasks (overkill).
• Real-time low-latency (slow).
• Cost-sensitive (multiple LLM calls).

5. Sample Multi-Agent — Content Production

Topic Input
    ↓
Researcher Agent (Claude Sonnet + Web Search)
    ↓ [Research Brief]
Writer Agent (Claude Opus, long form)
    ↓ [First Draft]
SEO Agent (specialized prompts)
    ↓ [Optimized Draft]
Editor Agent (Claude Sonnet, critic)
    ↓ [Final]
Publisher (no LLM, just publishes)

6. Cost Considerations

Single Agent Cost

• 1 LLM call per request.

Multi-Agent Cost

• 3-10x single agent cost.
• Each agent makes calls.
• Orchestration overhead.

When Worth It

• Quality matters more than cost.
• Saves human time significantly.
• Rare/important tasks.

7. Communication Patterns

Direct (Pipeline)

• Agent A output → Agent B input.

Shared Memory

• All agents read/write to common state.

Message Passing

• Agents communicate via messages.

8. Monitoring

• Each agent logged separately.
• Track time per agent.
• Track cost per agent.
• Identify bottleneck agents.

9. Common Pitfalls

❌ Too many agents — coordination cost > benefit. ❌ No clear roles — agents step on each other. ❌ Infinite loops between agents. ❌ Cost explosion — each agent calls LLM. ❌ No human override — wrong direction unstoppable.

10. Best Practices

• Start with 2-3 agents max.
• Clear role per agent (system prompts).
• Defined inputs/outputs per agent.
• Max iteration safety limit.
• Cost budget per task.
• Human-in-loop for critical.

11. Israel Specifics

• Hebrew multi-agent: works, more cost.
• Mixed teams (Hebrew + English) common.
• Israeli compliance: log agent decisions for audit.

12. אסיים בהמלצה.

פרומפט לדוגמה

Build research + writing + editing crew. CrewAI.

Multi-agent for code review. Architecture.

When NOT to use multi-agent?

© 2026 AI Expert Pro | גרסה 1.0.0