Menu

Invite me

LLM ≠ Agent

What actually turns a model into an agent.

Saai Syvendra · · CSE, University of Moratuwa

Open the slides Get the code

People use "LLM", "agent" and "harness" as if they mean the same thing. They don't. In this two-hour session we build one assistant, Campus Buddy, one piece at a time, and watch it go from a model that forgets your name to an agent that plans your week and writes the plan to a file.

An agent is an LLM, plus state, plus tools, running in a loop.

Recap

Go through the slides and you'll see the whole session comes down to one thing: know what an agent actually is before you hand it anything. The model is the brain. Everything else, the memory, the tools, the loop and the harness around it, is code you can write yourself. Once you have, you know exactly what you're handing over when you give an agent a job.

It's the same idea behind my one tip from HackerRank Orchestrate: if you don't understand the problem, neither will the AI agent.

Run it yourself

  1. Install Node.js 24 LTS (22 or later works).
  2. Get a free Gemini API key from Google AI Studio. Sign in with a Google account. No card needed.
  3. Clone the repo and install:
git clone https://github.com/saai-syvendra/llm-is-not-an-agent.git
cd llm-is-not-an-agent
npm install
  1. Copy .env.example to .env and paste your key after GEMINI_API_KEY=.
  2. Run the setup check. If it ends with 🎉, you're ready.
npm run doctor

If doctor fails, send me the error and we'll sort it out.

What we build

  1. LLM. A raw call to Gemini. Tell it your name, ask again, and it has no idea. A very smart function that forgets you instantly.
  2. Chatbot. Keep a list of messages and send the whole list every time. Now it remembers, because you remember for it.
  3. Agent. Give it a getWeather tool and write the loop yourself: run the tool it asks for, hand back the result, ask again, and stop when it stops asking.
  4. createAgent. The same agent in about ten lines, with a checkpointer for memory. You write a tool of your own.
  5. Harness. Deep Agents wraps that loop with a to-do list, a file system, skills and long-term memory. Ask it to plan your week before the mid-sem and it writes plan.md.

Steps 1, 2 and 5 are demos you can run along with. Steps 3 and 4 are yours to code.

Take-home challenge

  1. Give Campus Buddy a tool that matters to you: bus times from Katubedda, the canteen menu, your real timetable.
  2. Write a skill for something you know well, like how to prepare for one module's exam.
  3. Ask it to do a task that needs both, and watch which tool and which skill it reaches for.

Where to go next

  • The LangChain JS docs for agents, and Deep Agents for the harness.
  • Put a real UI on Campus Buddy instead of the terminal.
  • Read how coding agents like Claude Code and Cursor work. They run the same loop you wrote in step 3.
#Workshop#AI Agents#TypeScript#LangChain