Skip to content
Track Your DSA
Sign in

Agentic AI

Build agents by watching real ones work, and fail.

Every run here is a real model. Below, one searches seven times, finds nothing, and makes up an answer.

Start episode 1
Searching for something that isn't theregemini-3-flash-previewRecorded real run
model calls
1
tool calls
0
tokens in
75
tokens out
18
time
1.3s

Task What was the exact attendance at the first ever IPL match?

You
task
Model1.3s
search

Model call 1. It read 75 tokens and wrote 18. It asked for a tool, so it isn't done.

search({"query":"first ever IPL match attendance"})

stop_reason: tool_use

Context window75 tokens sent

  • System prompt19
  • Tool definitions42
  • Your messages14
step 1 of 15
New · Mission 1The ₹4.99 refund“Our bot refunded her ₹4.99 for a ₹499 cable, and told her it refunded ₹499.” Find it, fix it, prove it.

Act 0

Start here

  1. 1What is an AI agent?A chatbot guesses. An agent looks it up.4 min
  2. 2What is a language model?A machine that guesses what comes next, very well.4 min
  3. 3Prompts and messagesHow you talk to a model, and why the system prompt matters.4 min
  4. 4Your first API callGet a free key and talk to a model from code.10 min
  5. 5How this course worksWatch real runs, play the model, fix agents live, write code.4 min

Act 1

The model, as a builder sees it

  1. 6Tokens are the currencyWhat a model actually reads, and what you pay for.5 min
  2. 7Same question, different answersWhy a model can surprise you, even at temperature 0.5 min
  3. 8It remembers nothingEvery call starts from zero. Memory is your job.5 min
  4. 9Talk to code, not to peopleGetting output your program can trust.5 min
  5. 10The model asks, your code actsHow a model uses tools it can't touch.7 min

Act 2

The loop

  1. 11Twenty lines make an agentAsk, run, return, repeat: the code, running.6 min
  2. 12Many tools at onceOne reply, six tool calls.4 min
  3. 13Write the loopYour code, the real model's replies.15 min
  4. 14When agents go wrongThree real failures, and how to see them coming.10 min
  5. 15Fix it liveChange the prompt, run the real model, pass the check.15 min

Act 3

Tools and MCP

  1. 16The description is the contractA refund of ₹499 that came out as ₹4.99.8 min
  2. 17MCP: one plug for every toolThe standard way agents find and call tools, message by message.7 min
  3. 18Build an MCP serverBy hand first, then with the official SDK, plugged into Claude.20 min
  4. 19Your agent, real MCP toolsWhere the model, your agent and the server each do their part.6 min
  5. 20Errors the model can useWhen a tool fails, its message is the model's only clue.7 min

Act 4

Knowledge and memory

  1. 21Meaning as numbersHow “refund chahiye” finds the right English sentence.6 min
  2. 22Look it up firstSame question, three answers. Only one is right.6 min
  3. 23How you cut documents mattersSplit a rule in half and retrieval can't save you.5 min
  4. 24An agent that searchesIt decides what to look up, and when it has enough.15 min
  5. 25Memory across chatsRemembering a peanut allergy, days later.6 min
  1. Act 5Orchestrationcoming
  2. Act 6Trust: evals, guardrails, red teamingcoming
  3. Act 7Ship itcoming