Agent Fundamentals

What is an Agentic AI System?

In AI research, an agent is a system that perceives its environment, makes decisions, and takes actions to achieve a defined goal — often across multiple sequential steps without requiring human input at each stage. Unlike a standard chatbot that generates a single response and stops, an agentic system enters a continuous perceive → reason → act loop until the goal is satisfied or a stopping condition is reached.

Agentic systems can call external tools (search, code execution, APIs, databases), maintain memory across steps, and adapt their plan based on what they observe. This makes them capable of completing tasks that are too complex or multi-step for a single model inference.

Definition: What is an Agent?

An agent is an autonomous system that:

1. Observes the environment/current state
2. Reasons about how to achieve a goal
3. Acts by using tools or performing actions
4. Learns from outcomes to improve

Chatbot (reactive):
Input → Process → Output
(Stateless, responds to each query independently)

Agent (goal-driven):
Goal → Observe → Reason → Action → Observe → ... → Success
(Maintains state, tracks progress toward goal)

Agent vs Chatbot: Key Differences

Real-World Examples

Travel Agent

Goal: "Book a flight from NYC to LA for next Friday"

Steps:
1. Search flights (tool: flight API)
2. Check user preferences (tool: preferences database)
3. Book flight (tool: booking system)
4. Send confirmation (tool: email)

Result: Booking complete, email sent, calendar updated

Research Agent

Goal: "Find latest papers on diffusion models and summarize them"

Steps:
1. Search ArXiv (tool: arxiv API)
2. Download papers (tool: PDF downloader)
3. Extract text (tool: PDF parser)
4. Summarize each (LLM with reasoning)
5. Synthesize findings (LLM summary)

Result: Comprehensive research summary ready

Agent Capabilities

1. Tool Use

Tools available:
- Web search
- Calculator
- Code execution
- Database queries
- Email
- Calendar

Agent decides which tools to use based on goal!

2. Planning

Complex goal → Break into steps → Execute sequentially
"Write and publish a blog post" becomes:
- Write draft (LLM)
- Edit (human or agent)
- Format (tool)
- Publish (tool)
- Share (tool)

3. Reasoning

Agent thinks through approach:
"To book a flight, I need:
1. Dates (check calendar)
2. Preferences (check user DB)
3. Budget (check account)
4. Airlines (search)
5. Best option (compare & book)"

The Agent Loop (ReAct Pattern)

Thought: "I need to search for flights"
Action: search_flights(from="NYC", to="LA", date="2024-05-10")
Observation: [List of flights with prices]

Thought: "I should pick the cheapest direct flight"
Action: book_flight(flight_id=123, user_id=456)
Observation: "Booking successful, confirmation #789"

Thought: "I should send the user an email"
Action: send_email(to=user@example.com, subject="Flight Booked", body="...")
Observation: "Email sent"

Thought: "Goal accomplished!"

Why Agents are Revolutionary

Current AI limitations:
- LLMs can't access real-time info (knowledge cutoff)
- LLMs can't perform actions (just generate text)
- LLMs can't execute code reliably
- LLMs can't learn from environment

Agents solve this:
+ Can call APIs for real-time data
+ Can execute tools/code
+ Can interact with systems
+ Can improve via feedback loops