TodayTools

Agentic AI flashcards

AI that plans, uses tools and gets tasks done.

Donate
Beginner friendly, starts from zero
Section
How to study
0
Agentic AIStart from zero · 1 of 12

What is an AI , in one plain sentence?

An AI is a computer helper you give a goal to, and it works through the steps by itself, like searching, reading and writing, until the job is done.

A chatbot talks. An does.

All topics

Agentic AI at a glance

An AI agent is a computer helper that takes a goal and works through the steps itself until the job is done.

Like a new personal assistant: you give the goal, he runs the errands with the tools you allow, and asks before spending money.

Words you will meet

Agent
An AI helper that takes steps on its own toward a goal
Model
The AI brain that reads text and writes a reply
Token
A small piece of a word; models read and charge by tokens
Tool
One action the agent is allowed to use, like searching or reading a file
Loop
Think, act, look at the result, and repeat until done
Context
Everything the model can see right now, like papers on a desk
Trace
The written record of every step the agent took

The steps

  1. 1. Start from zero (12 cards): What an agent is, the words you need, and your first tiny tool.
  2. 2. Core basics (15 cards): What agents are, the loop, tools, reliability and human checks.
  3. 3. Next steps (16 cards): Tool calling, MCP, workflow patterns, limits and tracing.
  4. 4. Go deeper (16 cards): Injection, the lethal trifecta, sandboxing, evals, protocols and durability.
All 12 cards in Start from zero
What is an AI agent, in one plain sentence?
An AI agent is a computer helper you give a goal to, and it works through the steps by itself, like searching, reading and writing, until the job is done.
Before agents, what did you still have to do yourself?
An AI chat could tell you how to book a train ticket, but you still opened the site, filled the form and paid. An agent can do those steps for you, with your okay.
Where might you already meet an agent in daily life?
A coding helper that edits files and checks its own work, a support bot that really processes your refund, or an assistant that finds a free slot and books a meeting.
What everyday job is an agent most like?
A new personal assistant. You say 'book a table for four on Saturday'; he uses the phone and the diary, checks back with you, and asks before paying a deposit.
What is the 'model', and what is a 'token'?
The model is the AI brain: a program trained on huge amounts of text that reads your message and writes a reply. It reads and charges in tokens, small pieces of words.
What is a 'tool', a 'function' and an 'API'?
A tool is one action the agent may use, like 'search flights'. Behind it sits a function (a named piece of code that does one job), which often calls an API, a doorway one program uses to ask another for something.
What is the agent 'loop', and what is a 'step'?
The loop is think, act, look at the result, repeat. Each round is one step; a step limit caps the rounds so it cannot go on forever.
What is 'context', and how is it different from 'memory'?
Context is everything the model can see right now: your goal, the messages so far and tool results, like papers on a desk. Memory is notes saved in a file or database (an organised store of data) to read in a later session.
How do you know an agent really did the job? Learn: test, log, trace, eval.
A test is a small check that says pass or fail; a set of them is a test suite. A log or trace is the written record of every step, and an eval is a set of sample tasks used to score the agent.
What do you need before starting this topic?
Nothing to install for the ideas: just curiosity. For the code cards, reading a little Python helps, and running them needs an account with an AI provider and an API key (a secret password for your code).
Your first tiny win: what does this little tool do?
It is a tool with a name, a description the model reads to decide when to use it, and a result sent back. The model never runs it; your code does.
What will you be able to do after this topic?
Explain how an agent works, decide when a plain script is better, write the core loop with tools, and add limits and human checks so it stays safe.
Agentic AI cheatsheet

Building blocks

  • Model Decides the next action
  • Tools Typed functions your code runs
  • Loop Call, run tools, append results, repeat
  • Context Instructions, history, tool results, notes
  • Memory Files or a store it rereads across sessions
  • Stop rules Steps, time, tokens, money, repeat detection
  • Trace Every call, argument, result and cost

Patterns

  • Prompt chain Fixed steps, validate between them
  • Router Classify, then send to the right handler
  • Parallel Independent calls at once, then merge
  • Orchestrator-workers Lead splits, workers do, lead merges
  • Evaluator-optimiser Produce, critique, revise until pass
  • Autonomous agent Open-ended loop with tools and limits

Tool design

  • Name verb_noun, consistent prefixes
  • Description What, when to use, when not to
  • Schema Typed, described, required fields
  • Errors Actionable text, never silent empties
  • Output Trimmed, paginated, with ids
  • Idempotency key Safe retries for writes

MCP essentials

  • Server Exposes tools, resources, prompts
  • Client The AI app that connects to servers
  • stdio Local server as a child process
  • Streamable HTTP Remote server over HTTP
  • OAuth 2.1 Auth for remote servers
  • Elicitation Server asks the user for input mid-task

Safety checklist

  • Least privilege Read-only, scoped, short-lived tokens
  • Sandbox Container, no prod creds, egress allow-list
  • Break the trifecta No private data + untrusted input + outbound
  • Confirmation gate Code-enforced yes for irreversible acts
  • Budgets Step, token and spend caps per run
  • Pinned MCP servers Review tool definitions on update
Donate