Skill Agentic Frameworks — מסגרות סוכנים ל-Claude

מתי להשתמש

"AI Agent", "Autonomous AI", "Agentic", "LangChain", "CrewAI", "AutoGen".

הוראות עבודה

1. What is an Agent

AI that autonomously executes multi-step tasks using tools.

vs. Standard LLM

• LLM: input → output.
• Agent: input → think → use tool → think → use tool → ... → output.

2. Agent Building Blocks

א. Planning

• Break task into steps.
• "I need to: 1. Search, 2. Filter, 3. Email."

ב. Tool Use

• Search, APIs, calculators, code execution.
• See tool-use-function-calling skill.

ג. Memory

• Short-term: conversation context.
• Long-term: knowledge base, user preferences.

ד. Reflection

• Review own work.
• "Did I complete this correctly?"

ה. Loop

• Continue until task done.
• Stop conditions important.

3. Frameworks 2026

4. When to Use Agents

✅ Good Use Cases

• Multi-step research.
• Data gathering across sources.
• Complex customer support.
• Automated report generation.
• Coding (Claude Code is agent).

❌ Bad Use Cases

• Simple Q&A (no tools needed).
• Real-time critical (latency).
• Deterministic logic (use code).
• Risky automated decisions (financial, medical).

5. Sample Agent (Pseudo-code)

from anthropic import Anthropic

client = Anthropic()

tools = [
    {"name": "search_web", "description": "..."},
    {"name": "send_email", "description": "..."},
    {"name": "create_doc", "description": "..."},
]

messages = [{"role": "user", "content": "Research top 5 Israeli SaaS companies and email me a summary"}]

while True:
    response = client.messages.create(
        model="claude-sonnet-4-6",
        tools=tools,
        messages=messages
    )
    
    if response.stop_reason == "end_turn":
        break
    
    if response.stop_reason == "tool_use":
        tool_use = next(b for b in response.content if b.type == "tool_use")
        result = execute_tool(tool_use.name, tool_use.input)
        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": [{"type": "tool_result", "tool_use_id": tool_use.id, "content": result}]})

6. Production Considerations

Cost Control

• Max iterations — prevent runaway loops.
• Token budgets per task.
• Cache common operations.

Safety

• Approval before destructive actions (delete, send money).
• Sandbox for code execution.
• Rate limit API calls.

Reliability

• Idempotent tools — safe to retry.
• Fallback when stuck.
• Human-in-loop option.

7. Multi-Agent Systems

See multi-agent-systems skill.

8. Tool Use vs Agents

9. Israel Specifics

• Hebrew agents — Claude best.
• Local API integrations: WhatsApp Business, Israeli banks.
• Privacy: PII to agents — careful.

10. Common Pitfalls

❌ Infinite loops — set max iterations. ❌ Cost runaway — monitor. ❌ No human approval for destructive actions. ❌ Hallucinated tool calls — agent invents tools. ❌ Unclear task termination — when is "done"?

11. אסיים בהמלצה.

פרומפט לדוגמה

Build research agent that gathers data from 5 sources.

CrewAI vs LangChain — when each?

Production-ready agent for sales outreach. Architecture.

© 2026 AI Expert Pro | גרסה 1.0.0