AI Agents: From Chatbots to Autonomous Systems

The most overused word in 2025-2026 is “Agent.” But what is it?

Chatbot:

  • Input: “What is the weather in Tokyo?”
  • Output: “I don’t know, I don’t have internet access.” (Or it hallucinates).

Agent:

  • Input: “What is the weather in Tokyo?”
  • Action: Calls Weather API -> Gets data.
  • Output: “It is currently 12°C and raining in Tokyo.”

An Agent is an LLM with Tools and a Loop.

The Agent Loop (ReAct)

The core logic of an agent is often called ReAct (Reason + Act).

  1. Think: The user wants X. What do I need to do?
  2. Plan: I should use Tool A.
  3. Act: Execute Tool A.
  4. Observe: Look at the output of Tool A.
  5. Repeat: Is the task done? If no, go back to step 1.

Levels of Autonomy

Level 1: The Router

“If the user asks about billing, send to Stripe API. If tech support, send to Docs.” Simple, deterministic.

Level 2: The Tool User

“I have a calculator, a browser, and a file editor. I will figure out which one to use.” Flexible, but needs clear instructions.

Level 3: The Autonomous Agent

“Goal: ‘Make my website faster.’ Go.” The agent will:

  1. Browse the site.
  2. Run Lighthouse tests.
  3. SSH into the server.
  4. Edit the Nginx config.
  5. Restart the server.
  6. Verify results.

(We are currently wrestling with getting Level 3 to work reliably without destroying production.)

  • LangChain / LangGraph: The industry standard for building custom agent workflows.
  • Microsoft AutoGen: Multi-agent systems where “Manager Agent” talks to “Coder Agent” and “Reviewer Agent.”
  • CrewAI: A higher-level wrapper for role-playing agents.

The Challenge: Loops of Doom

The biggest risk with agents is infinite loops. Agent: “I need to fix the error.” Action: Edits file. Observation: “Error: Syntax error.” Agent: “I need to fix the error.” Action: Edits file (badly). Observation: “Error: Syntax error.” …repeats until your API credit card is maxed out.

Conclusion

Agents transform AI from a “Oracle” (knower of things) to a “Worker” (doer of things). They are harder to build, harder to debug, but infinitely more valuable.


Next: Tool Use and Function Calling — The mechanics of how agents work.