AI Agents Roadmap
This AI Agents Roadmap is a free, full-text syllabus of 101 lessons in 34 stages, starting with Basic Backend Development and ending with Safety + Red Team Testing. It suits anyone who wants to learn AI Agents, whether new to it or filling gaps. There is no fixed length: every lesson is open from the start, so you go at your own pace and skip what you already know.
Last updated
Reading is open to everyone. Sign in to join this roadmap and track what you finish.
Continue learning
1. Basic Backend DevelopmentBasic Backend Development
0%Basic Backend Development
1 topic · 0 completed
The diagram draws the same 101 lessons as a map, with lines showing the order. It needs JavaScript; the full syllabus below lists every lesson as text.
Full syllabus: every lesson in the AI Agents roadmap
34 stages and 101 lessons, in the order the roadmap suggests. Every lesson is free to read and open from the start.
Stage 1: Basic Backend Development
Basic Backend Development
Before you start learning how to build AI agents, we would recommend you to have a basic knowledge of Backend development. This includes, programming language knowledge, interacting with database and basics of APIs at minimum.
Stage 2: Git and Terminal Usage
Git and Terminal Usage
Git and the terminal are key tools for AI agents and developers. Git lets you track changes in code, work with branches, and collaborate safely with others. It stores snapshots of your work so you can undo mistakes or merge ideas. The terminal (command line) lets you move around files, run programs, set up servers,…
Stage 3: Streamed vs Unstreamed Responses
Streamed vs Unstreamed Responses
Streamed and unstreamed responses describe how an AI agent sends its answer to the user. With a streamed response, the agent starts sending words as soon as it generates them. The user sees the text grow on the screen in real time. This feels fast and lets the user stop or change the request early. It is useful for…
Stage 4: REST API Knowledge
REST API Knowledge
A REST API (Representational State Transfer) is an architectural style for designing networked applications. In AI agents, REST APIs enable communication between the agent and external systems, allowing for data exchange and integration. The agent can use REST APIs to retrieve data from external sources, send data…
Stage 5: Reasoning vs Standard Models
Reasoning vs Standard Models
Reasoning models break a task into clear steps and follow a line of logic, while standard models give an answer in one quick move. A reasoning model might write down short notes, check each note, and then combine them to reach the final reply. This helps it solve math problems, plan actions, and spot errors that…
Stage 6: Fine-tuning vs Prompt Engineering
Fine-tuning vs Prompt Engineering
Fine-tuning and prompt engineering are two ways to get better outputs from a language model. Fine-tuning means training an existing model further with your own examples so it adapts to specific tasks. It needs extra data, computing power, and time but creates deeply specialized models. Prompt engineering, in…
Stage 7: Transformer Models and LLMs
Transformer Models and LLMs
Transformer models are a type of neural network that read input data—like words in a sentence—all at once instead of one piece at a time. They use “attention” to find which parts of the input matter most for each other part. This lets them learn patterns in language very well. When a transformer has been trained on…
Stage 8: Embeddings and Vector Search
Embeddings and Vector Search
Embeddings turn words, pictures, or other data into lists of numbers called vectors. Each vector keeps the meaning of the original item. Things with similar meaning get vectors that sit close together in this number space. Vector search scans a large set of vectors and finds the ones nearest to a query vector, even…
Stage 10: Open Weight Models
Open Weight Models
Open-weight models are neural networks whose trained parameters, also called weights, are shared with everyone. Anyone can download the files, run the model, fine-tune it, or build tools on top of it. The licence that comes with the model spells out what you are allowed to do. Some licences are very permissive and…
Stage 9: Understand the Basics of RAG
Understand the Basics of RAG
RAG, short for Retrieval-Augmented Generation, is a way to make language models give better answers by letting them look things up before they reply. First, the system turns the user’s question into a search query and scans a knowledge source, such as a set of documents or a database. It then pulls back the most…
Stage 12: Closed Weight Models
Closed Weight Models
Closed-weight models are AI systems whose trained parameters—the numbers that hold what the model has learned—are not shared with the public. You can send prompts to these models through an online service or a software kit, but you cannot download the weights, inspect them, or fine-tune them on your own computer.…
Stage 11: Pricing of Common Models
Pricing of Common Models
When you use a large language model, you usually pay by the amount of text it reads and writes, counted in “tokens.” A token is about four characters or three-quarters of a word. Providers list a price per 1,000 tokens. For example, GPT-3.5 Turbo may cost around $0.002 per 1,000 tokens, while GPT-4 is much higher,…
Stage 13: Agent Loop
Agent Loop
An agent loop is the cycle that lets an AI agent keep working toward a goal. First, the agent gathers fresh data from its tools, sensors, or memory. Next, it updates its internal state and decides what to do, often by running a planning or reasoning step. Then it carries out the chosen action, such as calling an…
Perception / User Input
Perception, also called user input, is the first step in an agent loop. The agent listens and gathers data from the outside world. This data can be text typed by a user, spoken words, camera images, sensor readings, or web content pulled through an API. The goal is to turn raw signals into a clear, usable form. The…
Reason and Plan
Reason and Plan is the moment when an AI agent thinks before it acts. The agent starts with a goal and the facts it already knows. It looks at these facts and asks, “What do I need to do next to reach the goal?” It breaks the goal into smaller steps, checks if each step makes sense, and orders them in a clear path.…
Acting / Tool Invocation
Acting, also called tool invocation, is the step where the AI chooses a tool and runs it to get real-world data or to change something. The agent looks at its current goal and the plan it just made. It then picks the best tool, such as a web search, a database query, or a calculator. The agent fills in the needed…
Observation & Reflection
Observation and reflection form the thinking pause in an AI agent’s loop. First, the agent looks at the world around it, gathers fresh data, and sees what has changed. It then pauses to ask, “What does this new information mean for my goal?” During this short check, the agent updates its memory, spots errors, and…
Tokenization
Tokenization is the step where raw text is broken into small pieces called tokens, and each token is given a unique number. A token can be a whole word, part of a word, a punctuation mark, or even a space. The list of all possible tokens is the model’s vocabulary. Once text is turned into these numbered tokens, the…
Context Windows
A context window is the chunk of text a large language model can read at one time. It is measured in tokens, which are pieces of words. If a model has a 4,000-token window, it can only “look at” up to about 3,000 words before it must forget or shorten earlier parts. New tokens push old ones out, like a sliding…
Token Based Pricing
Token-based pricing is how many language-model services charge for use. A token is a small chunk of text, roughly four characters or part of a word. The service counts every token that goes into the model (your prompt) and every token that comes out (the reply). It then multiplies this total by a listed price per…
Temperature
Temperature is a setting that changes how random or predictable an AI model’s text output is. The value usually goes from 0 to 1, sometimes higher. A low temperature, close to 0, makes the model pick the most likely next word almost every time, so the answer is steady and safe but can feel dull or repetitive. A…
Top-p
Top-p, also called nucleus sampling, is a setting that guides how an LLM picks its next word. The model lists many possible words and sorts them by probability. It then finds the smallest group of top words whose combined chance adds up to the chosen p value, such as 0.9. Only words inside this group stay in the…
Frequency Penalty
Frequency penalty is a setting that tells a language model, “Stop repeating yourself.” As the model writes, it keeps track of how many times it has already used each word. A positive frequency-penalty value lowers the chance of picking a word again if it has been seen many times in the current reply. This helps cut…
Presence Penalty
Presence penalty is a setting you can adjust when you ask a large language model to write. It pushes the model to choose words it has not used yet. Each time a word has already appeared, the model gets a small score cut for picking it again. A higher penalty gives bigger cuts, so the model looks for new words and…
Stopping Criteria
Stopping criteria tell the language model when to stop writing more text. Without them, the model could keep adding words forever, waste time, or spill past the point we care about. Common rules include a maximum number of tokens, a special end-of-sequence token, or a custom string such as “\n\n”. We can also stop…
Max Length
Max Length sets the maximum number of tokens a language model can generate in one reply. Tokens are pieces of text—roughly 100 tokens equals a short paragraph. A small limit saves time and cost but risks cutting answers short. A large limit allows full, detailed replies but needs more compute and can lose focus.…
What are AI Agents?
An AI agent is a computer program or robot that can sense its surroundings, think about what it senses, and then act to reach a goal. It gathers data through cameras, microphones, or software inputs, decides what the data means using rules or learned patterns, and picks the best action to move closer to its goal.…
What are Tools?
Tools are extra skills or resources that an AI agent can call on to finish a job. A tool can be anything from a web search API to a calculator, a database, or a language-translation engine. The agent sends a request to the tool, gets the result, and then uses that result to move forward. Tools let a small core…
Stage 14: What is Prompt Engineering
What is Prompt Engineering
Prompt engineering is the skill of writing clear questions or instructions so that an AI system gives the answer you want. It means choosing the right words, adding enough detail, and giving examples when needed. A good prompt tells the AI what role to play, what style to use, and what facts to include or avoid. By…
Stage 15: Tree-of-Thought
Tree-of-Thought
Tree-of-Thought is a way to organize an AI agent’s reasoning as a branching tree. At the root, the agent states the main problem. Each branch is a small idea, step, or guess that could lead to a solution. The agent expands the most promising branches, checks if they make sense, and prunes paths that look wrong or…
Stage 16: Chain of Thought (CoT)
Chain of Thought (CoT)
Chain of Thought (CoT) is a way for an AI agent to think out loud. Before giving its final answer, the agent writes short notes that show each step it takes. These notes can list facts, name sub-tasks, or do small bits of math. By seeing the steps, the agent stays organized and is less likely to make a mistake.…
Personal assistant
A personal assistant AI agent is a smart program that helps one person manage daily tasks. It can check a calendar, set reminders, and send alerts so you never miss a meeting. It can read emails, highlight key points, and even draft quick replies. If you ask a question, it searches trusted sources and gives a short…
Code generation
Code-generation agents take a plain language request, understand the goal, and then write or edit source code to meet it. They can build small apps, add features, fix bugs, refactor old code, write tests, or translate code from one language to another. This saves time for developers, helps beginners learn, and…
Data analysis
AI agents can automate data analysis by pulling information from files, databases, or live streams. They clean the data by spotting missing values, outliers, and making smart corrections. After cleaning, agents find patterns like sales spikes or sensor drops and can build charts or dashboards. Some run basic…
Web Scraping / Crawling
Web scraping and crawling let an AI agent collect data from many web pages without human help. The agent sends a request to a page, reads the HTML, and pulls out parts you ask for, such as prices, news headlines, or product details. It can then follow links on the page to reach more pages and repeat the same steps.…
NPC / Game AI
Game studios use AI agents to control non-player characters (NPCs). The agent observes the game state and decides actions like moving, speaking, or fighting. It can shift tactics when the player changes strategy, keeping battles fresh instead of predictable. A quest giver might use an agent to offer hints that fit…
Stage 17: Tool Definition
Tool Definition
A tool is any skill or function that an AI agent can call to get a job done. It can be as simple as a calculator for math or as complex as an API that fetches live weather data. Each tool has a name, a short description of what it does, and a clear list of the inputs it needs and the outputs it returns. The agent’s…
Be specific in what you want
When you ask an AI to do something, clear and exact words help it give the answer you want. State the goal, the format, and any limits up front. Say who the answer is for, how long it should be, and what to leave out. If numbers, dates, or sources matter, name them. For example, rather than “Explain World War II,”…
Provide additional context
Provide additional context means giving the AI enough background facts, constraints, and goals so it can reply in the way you need. Start by naming the topic and the purpose of the answer. Add who the answer is for, the tone you want, and any limits such as length, format, or style. List key facts, data, or…
Use relevant technical terms
When a task involves a special field such as law, medicine, or computer science, include the correct domain words in your prompt so the AI knows exactly what you mean. Ask for “O(n log n) sorting algorithms” instead of just “fast sorts,” or “HTTP status code 404” instead of “page not found error.” The right term…
Use Examples in your Prompt
A clear way to guide an AI is to place one or two short samples inside your prompt. Show a small input and the exact output you expect. The AI studies these pairs and copies their pattern. Use plain words in the sample, keep the format steady, and label each part so the model knows which is which. If you need a…
Iterate and Test your Prompts
After you write a first prompt, treat it as a draft, not the final version. Run it with the AI, check the output, and note what is missing, wrong, or confusing. Change one thing at a time, such as adding an example, a limit on length, or a tone request. Test again and see if the result gets closer to what you want.…
Specify Length, format etc
When you give a task to an AI, make clear how long the answer should be and what shape it must take. Say “Write 120 words” or “Give the steps as a numbered list.” If you need a table, state the column names and order. If you want bullet points, mention that. Telling the AI to use plain text, JSON, or markdown stops…
Stage 18: Model Context Protocol (MCP)
Model Context Protocol (MCP)
Model Context Protocol (MCP) is an open, standardized protocol that lets AI applications connect to external data sources and tools in a uniform way. Instead of writing a custom integration every time a model needs to read a file, query a database, or call an API, developers expose that capability through an MCP…
Web Search
Web search lets an AI agent pull fresh facts, news, and examples from the internet while it is working. The agent turns a user request into search words, sends them to a search engine, and reads the list of results. It then follows the most promising links, grabs the page text, and picks out the parts that answer…
Code Execution / REPL
Code Execution or REPL (Read-Eval-Print Loop) lets an AI agent run small pieces of code on demand, see the result right away, and use that result to decide what to do next. The agent “reads” the code, “evaluates” it in a safe sandbox, “prints” the output, and then loops back for more input. With this tool the agent…
Database Queries
Database queries let an AI agent fetch, add, change, or remove data stored in a database. The agent sends a request written in a query language, most often SQL. The database engine then looks through its tables and returns only the rows and columns that match the rules in the request. With this tool, the agent can…
API Requests
API requests let an AI agent ask another service for data or for an action. The agent builds a short message that follows the service’s rules, sends it over the internet, and waits for a reply. For example, it can call a weather API to get today’s forecast or a payment API to charge a customer. Each request has a…
Email / Slack / SMS
Email, Slack, and SMS are message channels an AI agent can use to act on tasks and share updates. The agent writes and sends emails to give detailed reports or collect files. It posts to Slack to chat with a team, answer questions, or trigger alerts inside a workspace. It sends SMS texts for quick notices such as…
File System Access
File system access lets an AI agent read, create, change, or delete files and folders on a computer or server. With this power, the agent can open a text file to pull data, write a new report, save logs, or tidy up old files without human help. It can also move files between folders to keep things organized. This…
Stage 19: Creating MCP Servers
Creating MCP Servers
An MCP server stores and shares conversation data for AI agents using the Model Context Protocol (MCP), a standard for agent memory management. Start by picking a language and web framework, then create REST endpoints like /messages, /state, and /health. Each endpoint exchanges JSON following the MCP schema. Store…
Stage 20: What is Agent Memory?
What is Agent Memory?
Agent memory is the part of an AI agent that keeps track of what has already happened. It stores past user messages, facts the agent has learned, and its own previous steps. This helps the agent remember goals, user likes and dislikes, and important details across turns or sessions. Memory can be short-term,…
Stage 21: Short Term Memory
Short Term Memory
Short term memory are the facts which are passed as a part of the prompt to the LLM e.g. there might be a prompt like below:
MCP Hosts
MCP Hosts are computers or services that run the Model Context Protocol. They handle incoming calls, load the MCP manifest, check requests, and pass data between users, tools, and language models. Hosts may cache recent messages, track token usage, and add safety or billing checks before sending prompts to the…
MCP Client
The MCP Client is the part of an AI agent that talks to the language model API. It collects messages, files, and tool signals, packs them using the Model Context Protocol, and sends them to the model. When a reply comes back, it unpacks it, checks the format, and passes the result to other modules. It also tracks…
MCP Servers
An MCP Server is the main machine or cloud service that runs the Model Context Protocol. It keeps the shared “memory” that different AI agents need so they stay on the same page. When an agent sends a request, the server checks who is asking, pulls the right context from its store, and sends it back fast. It also…
Local Desktop
A Local Desktop deployment means running the MCP server directly on your own computer instead of a remote cloud or server. You install the MCP software, needed runtimes, and model files onto your desktop or laptop. The server then listens on a local address like 127.0.0.1:8000, accessible only from the same machine…
Remote / Cloud
Remote or cloud deployment places the MCP server on a cloud provider instead of a local machine. You package the server as a container or virtual machine, choose a service like AWS, Azure, or GCP, and give it compute, storage, and a public HTTPS address. A load balancer spreads traffic, while auto-scaling adds or…
Stage 22: Episodic vs Semantic Memory
Episodic vs Semantic Memory
Agent memory often has two parts. Episodic memory is relevant to the context of the current conversation and may be lost after the conversation ends. Semantic memory is relevant to the broader knowledge of the agent and is persistent.
Stage 23: Long Term Memory
Long Term Memory
Long term memory in an AI agent stores important information for future use, like a digital notebook. It saves facts, past events, user preferences, and learned skills so the agent can make smarter and more consistent decisions over time. Unlike short-term memory, this data survives across sessions. When a similar…
Stage 24: RAG Agent
RAG Agent
A RAG (Retrieval-Augmented Generation) agent mixes search with language generation so it can answer questions using fresh and reliable facts. When a user sends a query, the agent first turns that query into an embedding—basically a number list that captures its meaning. It then looks up similar embeddings in a…
RAG and Vector Databases
RAG, short for Retrieval-Augmented Generation, lets an AI agent pull facts from stored data each time it answers. The data sits in a vector database. In that database, every text chunk is turned into a number list called a vector. Similar ideas create vectors that lie close together, so the agent can find related…
User Profile Storage
User profile storage is the part of an AI agent’s memory that holds stable facts about each user, such as name, age group, language, past choices, and long-term goals. The agent saves this data in a file or small database so it can load it each time the same user returns. By keeping the profile separate from…
Summarization / Compression
Summarization or compression lets an AI agent keep the gist of past chats without saving every line. After a talk, the agent runs a small model or rule set that pulls out key facts, goals, and feelings and writes them in a short note. This note goes into long-term memory, while the full chat can be dropped or…
Forgetting / Aging Strategies
Forgetting or aging strategies help an AI agent keep only the useful parts of its memory and drop the rest over time. The agent may tag each memory with a time stamp and lower its importance as it gets older, or it may remove items that have not been used for a while, much like a “least-recently-used” list. Some…
Stage 25: ReAct (Reason + Act)
ReAct (Reason + Act)
ReAct is an agent pattern that makes a model alternate between two simple steps: Reason and Act. First, the agent writes a short thought that sums up what it knows and what it should try next. Then it performs an action such as calling an API, running code, or searching a document. The result of that action is fed…
Stage 26: Planner Executor
Planner Executor
A planner-executor agent is a type of AI agent that splits its work into two clear parts: planning and execution. The planner thinks ahead, taking a goal and breaking it down into a sequence of steps, ordering them in a logical and efficient manner. The executor, on the other hand, takes each planned step and…
Stage 27: DAG Agents
DAG Agents
A DAG (Directed Acyclic Graph) agent is made of small parts called nodes that form a one-way graph with no loops. Each node does a task and passes its result to the next. Because there are no cycles, data always moves forward, making workflows easy to follow and debug. Independent nodes can run in parallel,…
Stage 28: Manual (from scratch)
Manual (from scratch)
Building an AI agent from scratch means writing every part of the system yourself, without ready-made libraries. You define how the agent senses inputs, stores memory, makes decisions, and learns over time. First, you pick a clear goal, like solving puzzles or chatting. Then you code the inputs (keyboard, mouse,…
Stage 29: LLM Native "Function Calling"
LLM Native "Function Calling"
LLM native “function calling” lets a large language model decide when to run a piece of code and which inputs to pass to it. You first tell the model what functions are available. For each one you give a short name, a short description, and a list of arguments with their types. During a chat, the model can answer…
OpenAI Functions Calling
OpenAI Function Calling lets you give a language model a list of tools and have it decide which one to use and with what data. You describe each tool with a short name, what it does, and the shape of its inputs in a small JSON-like schema. You then pass the user message and this tool list to the model. Instead of…
OpenAI Assistant API
The OpenAI Assistants API lets you add clear, task-specific actions to a chat with a large language model. You first describe each action you want the model to use, giving it a name, a short purpose, and a list of inputs in JSON form. During the chat, the model may decide that one of these actions will help. It…
Gemini Function Calling
Gemini function calling lets you hook the Gemini language model to real code in a safe and simple way. You first list the functions you want it to use, each with a name, a short note about what it does, and a JSON schema for the needed arguments. When the user speaks, Gemini checks this list and, if a match makes…
Anthropic Tool Use
Anthropic Tool Use lets you connect a Claude model to real software functions so the agent can do useful tasks on its own. You give Claude a list of tools, each with a name, a short description, and a strict JSON schema that shows the allowed input fields. During a chat you send user text plus this tool list.…
Multi-Agents
Multi-agent systems involve multiple autonomous agents that interact with each other to achieve individual or collective goals. These agents can collaborate, compete, or coordinate their actions within a shared environment. The interactions between these agents can lead to emergent behaviors and solutions that are…
Self-critique Agents
Self-critique agents are a type of AI agent designed to evaluate and improve their own performance. They work by generating outputs, then analyzing those outputs to identify weaknesses or errors. Based on this self-analysis, the agent refines its approach and attempts to produce better results in subsequent…
Metrics to Track
To judge how well an AI agent works, you need clear numbers. Track accuracy, precision, recall, and F1 score to measure correctness. For ranking tasks, use metrics like mean average precision or ROC-AUC. If users interact with the agent, monitor response time, latency, and failure rates. Safety metrics count toxic…
Unit Testing for Individual Tools
Unit testing checks that each tool an AI agent uses works as expected when it stands alone. You write small tests that feed the tool clear input and then compare its output to a known correct answer. If the tool is a function that parses dates, you test many date strings and see if the function gives the right…
Integration Testing for Flows
Integration testing for flows checks that an AI agent works well from the first user input to the final action, across every step in between. It joins all parts of the system—natural-language understanding, planning, memory, tools, and output—and runs them together in real scenarios. Test cases follow common and…
Human in the Loop Evaluation
Human-in-the-loop evaluation checks an AI agent by letting real people judge its output and behavior. Instead of trusting only automated scores, testers invite users, domain experts, or crowd workers to watch tasks, label answers, flag errors, and rate clarity, fairness, or safety. Their feedback shows problems…
Stage 30: Prompt Injection / Jailbreaks
Prompt Injection / Jailbreaks
Prompt injection, also called a jailbreak, is a trick that makes an AI system break its own rules. An attacker hides special words or symbols inside normal-looking text. When the AI reads this text, it follows the hidden instructions instead of its safety rules. The attacker might force the AI to reveal private…
LangSmith
LangSmith is a tool that helps you see how well your AI agents work. It lets you record every step the agent takes, from the first input to the final answer. You can replay these steps to find places where the agent goes wrong. LangSmith also lets you create test sets with real user prompts and compare new model…
DeepEval
DeepEval is an open-source tool that helps you test and score the answers your AI agent gives. You write small test cases that show an input and the reply you hope to get, or a rule the reply must follow. DeepEval runs the agent, checks the reply with built-in measures such as similarity, accuracy, or safety, and…
Ragas
Ragas is an open-source tool used to check how well a Retrieval-Augmented Generation (RAG) agent works. You give it the user question, the passages the agent pulled from a knowledge base, and the final answer. Ragas then scores the answer for things like correctness, relevance, and whether the cited passages really…
LangChain
LangChain is a framework designed to simplify the creation of applications using large language models (LLMs). It provides tools and abstractions to connect LLMs to various data sources, create chains of calls to LLMs or other utilities, and build agents that can interact with their environment. Essentially, it…
LangGraph
LangGraph is a Python library designed to help developers create robust and stateful conversational AI applications, often referred to as AI Agents. It allows you to structure complex agent workflows by defining states (data held throughout the conversation) and edges (transitions between states based on agent…
Haystack
Haystack is an open-source Python framework that helps you build search and question-answering agents fast. You connect your data sources, pick a language model, and set up pipelines that find the best answer to a user’s query. Haystack handles tasks such as indexing documents, retrieving passages, running the…
LlamaIndex
LlamaIndex is an open-source Python toolkit that helps you give a language model access to your own data. You load files such as PDFs, web pages, or database rows. The toolkit breaks the text into chunks, turns them into vectors, and stores them in a chosen vector store like FAISS or Pinecone. When a user asks a…
CrewAI
CrewAI is an open-source Python framework for creating teams of AI agents, called a crew. Each agent is assigned a name, role, and set of tools, and the system manages planning, communication, and execution between them. To use it, install the package, define agents in code, connect them with a Crew object, and…
AutoGen
AutoGen is an open-source Python framework that helps you build AI agents without starting from scratch. It lets you define each agent with a role, goals, and tools, then handles the chat flow between them and a large language model such as GPT-4. You can chain several agents so they plan, code, review, and run…
Smol Depot
Smol Depot is an open-source kit that lets you bundle all the parts of a small AI agent in one place. You keep prompts, settings, and code files together in a single folder, then point the Depot tool at that folder to spin the agent up. The tool handles tasks such as loading models, saving chat history, and calling…
Agno
Agno is a Python framework designed to streamline the process of building AI agents. It offers tools and abstractions that simplify tasks such as agent planning, tool use, and memory management, making it easier to create sophisticated and functional AI agents without managing low-level implementation details. It…
Structured logging & tracing
Structured logging and tracing are ways to record what an AI agent does so you can find and fix problems fast. Instead of dumping plain text, the agent writes logs in a fixed key-value format, such as time, user\id, step, and message. Because every entry follows the same shape, search tools can filter, sort, and…
LangSmith
LangSmith is a web tool that helps you see and fix what your AI agents are doing. It records each call that the agent makes to a language model, the input it used, and the answer it got back. You can replay any step, compare different prompts, measure cost, speed, and error rates, and tag runs for easy search. It…
Helicone
Helicone is an open-source tool that helps you watch and understand how your AI agents talk to large language models. You send your model calls through Helicone’s proxy, and it records each request and response without changing the result. A clear web dashboard then shows logs, latency, token counts, error rates,…
LangFuse
LangFuse is a free, open-source tool that lets you watch and debug AI agents while they run. You add a small code snippet to your agent, and LangFuse starts collecting every prompt, model response, and user input. It shows this data as neat timelines, so you can see each step the agent takes, how long the calls…
openllmetry
openllmetry is a small Python library that makes it easy to watch what your AI agent is doing and how well it is working. It wraps calls to large-language-model APIs, vector stores, and other tools, then sends logs, traces, and simple metrics to any backend that speaks the OpenTelemetry standard, such as Jaeger,…
Stage 31: Tool sandboxing / Permissioning
Tool sandboxing / Permissioning
Tool sandboxing keeps the AI agent inside a safe zone where it can only run approved actions and cannot touch the wider system. Permissioning sets clear rules that say which files, networks, or commands the agent may use. Together they stop errors, leaks, or abuse by limiting what the agent can reach and do.…
Stage 32: Data Privacy + PII Redaction
Data Privacy + PII Redaction
AI agents often process text, images, and logs that include personal data like names, phone numbers, or addresses. Leaks can cause fraud, stalking, or other harm, so laws like GDPR and CCPA require strict protections. A key method is PII redaction: scanning inputs and outputs to find and mask any personal details…
Stage 33: Bias & Toxicity Guardrails
Bias & Toxicity Guardrails
Bias and toxicity guardrails keep an AI agent from giving unfair or harmful results. Bias shows up when training data favors certain groups or views. Toxicity is language that is hateful, violent, or rude. To stop this, start with clean and balanced data. Remove slurs, stereotypes, and spam. Add examples from many…
Stage 34: Safety + Red Team Testing
Safety + Red Team Testing
Safety + Red Team Testing is the practice of checking an AI agent for harmful or risky behavior before and after release. Safety work sets rules, guardrails, and alarms so the agent follows laws, keeps data private, and treats people fairly. Red team testing sends skilled testers to act like attackers or…
Questions about the AI Agents Roadmap
Is the AI Agents Roadmap free?
Yes. All 101 lessons in this roadmap are free to read on Bibha AI Learning Labs, with no account and no payment. Signing in is optional: it lets you mark lessons as done and keep your progress, and it opens the AI tutor, which runs on your own API key.
How long does the AI Agents Roadmap take?
There is no fixed duration. It has 101 lessons in 34 stages, and the time it takes depends on how many hours you give it each week. Nothing is locked, so you can skip what you already know, and your progress is saved once you sign in.
Where should I start the AI Agents Roadmap?
Start with stage 1, Basic Backend Development, and work down in order. The stages are numbered in the order the roadmap suggests, but every lesson is open, so if you already know Basic Backend Development you can go straight to Git and Terminal Usage or anything after it. Each lesson links to free resources for going deeper.
Can I learn AI Agents in Hindi?
The lessons are written in English. Once you sign in, you can ask the AI tutor on any lesson to explain it in Hindi or Hinglish. It answers from that lesson’s own text, so the explanation stays on the syllabus. The tutor runs on your own Google, OpenAI or Anthropic key.