AI Agents & Automation

AI agents represent a paradigm shift from simple chatbots to autonomous systems that can perceive their environment, reason about what to do, take actions using tools, and iterate until a task is complete.

---

What Is an AI Agent?

The Agent Loop:
===============

                +-------------+
                |   Observe   |
                | Environment |
                +------+------+
                       |
                       v
                +-------------+
          +---->|   Reason    |
          |     |  & Plan     |
          |     +------+------+
          |            |
          |            v
          |     +-------------+
          |     |    Act      |
          |     | (Use Tools) |
          |     +------+------+
          |            |
          |            v
          |     +-------------+
          |     |  Evaluate   |------ Goal met? ---> DONE
          |     |  Result     |
          |     +------+------+
          |            |
          |         No |
          +------------+
             (iterate)

Key insight: An agent takes multiple steps autonomously rather than generating a single response.

---

Tool Use and Function Calling

Common Agent Tools

Tool Category	Examples	What It Enables
Code execution	Python interpreter, shell	Run calculations, process data
File operations	Read, write, edit files	Manage documents and code
Web browsing	Search, web scraping	Access current information
API calls	REST APIs, databases	Interact with services
Communication	Email, Slack	Collaborate with humans

Function Calling Flow

User: "What's the weather in London right now?"

LLM outputs:  {
                "tool": "get_weather",
                "arguments": {"location": "London, UK", "units": "metric"}
              }

System executes: get_weather("London, UK", "metric")
Returns: {"temp": 12, "condition": "cloudy", "humidity": 78}

LLM: "It is currently 12 degrees C and cloudy in London, with
      78% humidity."

The model does not execute functions directly — it outputs structured requests, and application code handles execution. This keeps the system secure and controllable.

---

Multi-Step Reasoning and Planning

Multi-Step Agent Example:
Goal: "Analyse Q3 sales data and create a summary report"

Step 1:  [Think] Find the Q3 sales data file
Step 2:  [Act]   Search for files matching "Q3*sales*"
Step 3:  [Think] Found "Q3_sales_2025.csv" -- read it
Step 4:  [Act]   Read the CSV file
Step 5:  [Think] Analyse trends, top products, anomalies
Step 6:  [Act]   Run Python code for statistics
Step 7:  [Think] Create visualisations
Step 8:  [Act]   Generate charts with matplotlib
Step 9:  [Think] Compile into a report
Step 10: [Act]   Write the report to a file
Step 11: [Done]  Report created

The ReAct Pattern

ReAct (Reasoning + Acting) interleaves thinking and acting:

Thought: I need the population of Tokyo.
Action: web_search("population of Tokyo 2025")
Observation: Approximately 13.96 million (city) or 37.4 million (metro).
Thought: The user probably wants the metro figure.
Answer: Tokyo has about 14 million in the city proper and
        37.4 million in the Greater Tokyo Area.

---

Agent Architectures

Single Agent

One LLM handles all reasoning and tool use. Simple but limited for complex tasks.

Multi-Agent Systems