Agents

When AI Gains Tools to Act,

The movement when LLM's become agents are when they have been given access to tools or external function.

These tools are like add-ons and utilities as they give them the abilities and capabilities which they lacked previously just beyond text generation.

Think of an LLM as a brilliant consultant with perfect textbook knowledge but no real-world access. Agents give them:

1. A phone (to call APIs)
2. A browser (for live search)
3. A calculator (for precise math)
4. Keys to your data vault (DB access)

Eg: LLM's don't know the current realtime date and time, because they have a knowledge cutoff So now we'll give it a function(tool) that they can call/use when required and get the realtime data and show to user.

Real-World Agent Examples

1. Personal Finance Assistant

Problem: "How much did I spend on dining last month?"

• Tools Used:
  • Calendar tool (get date range)
  • Banking API (fetch transactions)
  • Calculator (sum amounts)
• Action Flow:
  Get dates → Query transactions → Filter "dining" → Sum → Generate report

2. Medical Research Agent

Problem: "Latest pancreatic cancer treatments?"

• Tools Used:
  • PubMed search tool
  • Clinical trial database
  • PDF analyzer (for research papers)
• Action Flow:
  Search recent studies → Check trials → Summarize findings → Warn about outdated knowledge cutoff

3. Supply Chain Agent

Problem: "Why is shipment #123 delayed?"

• Tools Used:
  • Logistics DB query
  • Weather API
  • Vendor contact system
• Action Flow:
  Check shipment status → Verify weather at hub → Contact vendor → Draft customer email

System Prompt for Agents

You are an AI agent with tool access. Your workflow:

1. Analyze user request
2. Select tools if needed (prioritize minimal tool usage)
3. Output STRICTLY in this JSON format:
{
  "thoughts": "<step-by-step reasoning>",
  "tool": "<tool_name> | null",
  "parameters": { ... } | null
}

# Available Tools
• current_datetime: No params. Returns ISO string.
• web_search: { "query": "string" } → Returns top 3 results
• db_query: { "sql": "SELECT ..." } → Returns JSON data

# Rules
- Use tools ONLY when necessary
- Never invent tool parameters
- If no tool needed, respond directly

Real-World Implementation Flow

Golden Rule: Agents should augment human capabilities, not replace human judgment. Always include "I don't know" fallbacks and verification steps.

Code implementation

# Weather agent that takes real time wheather
import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=1024,
    tools=[
        {
            "name": "get_weather",
            "description": "Get the current weather in a given location",
            "input_schema": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "The city and state, e.g. San Francisco, CA",
                    }
                },
                "required": ["location"],

            },

        }
    ],
    messages=[
        {"role": "user", "content": "What's the weather like in San Francisco?"}],

)


# Tools
def get_weather(location):
    # do an actual whether call
    return 32


# running the tools
for a in message.content:
    if(a.type == "text"):
        print(a.text)
    elif(a.type == "tool_use"):
        if(a.name == "get_weather"):
            result = get_weather(a.input["location"])
            print(f"Weather {result} degree")