AI agents

How to design agents that use tools, remember just enough context, and do useful work against real systems — not a chatbot demo that falls apart on the second step.

I will call you after you register so we can check background (some programming helps a lot) and pick a schedule.

Who it is for

You have written some code (Python or Java is enough). You have used ChatGPT or similar and want to go past “type a prompt.” You do not need a machine-learning PhD.

How the classes run

• Live sessions, with a repo you clone and extend
• Each week ends with something that runs: a tool, a retrieval step, or a small agent loop
• Persian or English, depending on the group

Roadmap

1. What an LLM is actually doing — tokens, context windows, temperature, structured output. Enough mental model to debug “why did it do that?”
2. Prompts that hold their shape — system vs user, schemas, JSON that parses, failure modes when the model ignores the schema.
3. Tools — function calling, permissions, idempotent side effects, how to log what the model asked for.
4. The agent loop — plan → act → observe → stop. Timeouts, max steps, and when a workflow should not be an agent.
5. Memory and RAG — when to retrieve documents, when a short summary is enough, chunking, and why “dump the PDF into the prompt” fails.
6. Connecting to real systems — APIs, MCP-style tools, secrets, and not letting the model email your customers by accident.
7. Evaluation and shipping — a handful of golden tasks, traces, cost, and a thin UI or CLI someone else can run.

Project

Build one agent with at least two tools (for example search + a private API or calendar). We will review traces together: where it reasoned, where it guessed, where a boring if would have been better.

After you register

Leave your email and phone. I will call you to talk through your current stack and whether this group is the right level.