$ roadmap --ai-agents
Agentic AI & AI Agents Roadmap for Beginners
The skill-by-skill agentic AI roadmap for becoming someone who can build AI agents — each stage with what to learn, a project that proves it, and a checkpoint that tells you you're ready to move on. If you are starting from zero with AI in general, begin with the broader guide on how to learn AI.
By Abdul Rahman Azam, founder of AI Season · Updated
Key facts
| Starting point | Basic Python, Git and calling a web API |
|---|---|
| Stages | LLM APIs → prompting → tools and the agent loop → RAG → frameworks and patterns → memory → MCP → evaluation → security → deployment → specialise |
| Projects | One small, shippable project per stage — a portfolio of 8–10 by the end |
| Time | About 3–4 months part-time; a live bootcamp covers the core in 6 weeks |
| Hardware | Any laptop — models run on provider APIs or small local models |
Stage 0 — Prerequisites
Be comfortable with Python functions, lists and dictionaries, reading and writing JSON, calling an HTTP API, virtual environments and Git.
Checkpoint: you can write a script that fetches data from a public API, transforms it and saves the result — without copying a tutorial line by line.
Stage 1 — LLM fundamentals and APIs
Understand tokens, context windows, temperature and why models are stateless. Call a model from code, stream the output, keep a conversation history and handle rate limits and errors.
Project: a command-line chat app with streaming, history and a system prompt you can switch.
Stage 2 — Prompting and structured output
Write system prompts with a role, rules and an output format; use few-shot examples; request JSON that matches a schema and validate it in code.
Project: a CV or invoice extractor that returns validated JSON and rejects malformed output.
Stage 3 — Tool calling and the agent loop
Describe functions as tools, let the model request them, run them safely, return results and loop until done — with a step limit and an allow-list of tools.
Project: follow the build-an-agent tutorial, then extend it with a tool of your own.
Checkpoint: you can explain, without notes, what happens between the user's message and the agent's final answer.
Stage 4 — RAG: your own data
Chunking, embeddings, vector databases, hybrid search, reranking, citations and honest refusals. Measure retrieval separately from answers. See RAG for AI agents.
Project: a Q&A agent over a real document set — your university handbook or a company policy — that cites its sources and says when it doesn't know.
Stage 5 — Agent patterns and frameworks
Learn the patterns: ReAct, planning, reflection, routing, and supervisor-with-specialists. Then learn LangChain components and LangGraph for state, branching, retries and human approval. Compare options in AI agent frameworks compared.
Project: a LangGraph workflow that plans a multi-step task, retries a failing step and pauses for approval before a risky action.
Stage 6 — Memory and state
Short-term memory (the conversation), trimming and summarising long histories, long-term memory stored in a database or vector store, and per-user state.
Project: an assistant that remembers a user's preferences across sessions — and forgets on request.
Stage 7 — MCP and integrations
Use the Model Context Protocol to connect agents to tools and data in a standard way; write a small MCP server; understand where agent-to-agent protocols such as A2A fit. See the MCP guide.
Project: an MCP server that exposes one real tool (a database query or a file search) and an agent that uses it.
Stage 8 — Evaluation and observability
Build test sets, automate checks, trace every model and tool call, and compare versions before shipping. This is the skill that separates demos from products.
Project: add a 30-question evaluation suite and tracing to your RAG agent; improve its score and write down what changed.
Stage 9 — Security and guardrails
Prompt injection, data leaks, unsafe tool calls, permission design, sandboxes, cost and rate limits, and human approval for irreversible actions.
Project: attack your own agent with injected instructions in a document, then add the guardrails that stop it.
Stage 10 — Deployment and cost
Put the agent behind an API, manage secrets and configuration, deploy to a public URL, add caching, choose cheaper models where quality allows, and monitor latency and spend.
Project: a deployed agent with a small web UI or a WhatsApp front end that a friend can actually use.
Stage 11 — Specialise
- Voice agents — speech-to-text, turn-taking, latency and text-to-speech.
- Multimodal agents — images, documents, audio and video as inputs.
- Multi-agent systems — specialists, handoffs and agent-to-agent protocols.
- Coding agents — agents that edit code and run tests until they pass.
- Browser and computer-use agents — operating websites and apps like a person.
Suggested timeline
| Weeks (part-time) | Stages | Milestone |
|---|---|---|
| 1–2 | 0–2 | Chat app and structured extractor on GitHub |
| 3–4 | 3 | First tool-using agent |
| 5–6 | 4 | Document Q&A agent with citations |
| 7–9 | 5–6 | LangGraph workflow with memory and approval |
| 10–11 | 7–9 | MCP server, evaluation suite and guardrails |
| 12–14 | 10–11 | Deployed agent and one specialisation |
The AI Season bootcamp covers stages 1–10 plus voice agents live in 12 sessions over 6 weeks — explained in Urdu, coded in English — with a Mid Test, a Final Test and a top-3 refund. Cohort 02 starts 1st January 2027.
Frequently asked questions
How long does it take to learn to build AI agents?
With basic Python, about three to four months of steady part-time work to finish this roadmap, or 6 weeks for the core in a structured live bootcamp.
What should I learn first for AI agents?
Python basics, then calling an LLM API, then tool calling and the agent loop. RAG and frameworks come after you have built one agent by hand.
Do I need machine learning for AI agents?
No. Agent engineering is built on hosted models. Programming, prompting, retrieval, tool design, evaluation and deployment matter far more than ML theory.
Which projects should be in an AI agents portfolio?
A document Q&A agent with citations, a tool-using agent, a LangGraph workflow with approval steps, an MCP server, and one deployed agent people can use — each with a short write-up of how you tested it.
Is an agentic AI roadmap different from an AI agents roadmap?
No. Agentic AI is the umbrella term for AI that plans and acts towards a goal, and AI agents are the systems you build to do it, so both roadmaps cover the same skills: LLM APIs, tool calling, RAG, frameworks such as LangGraph, MCP, memory, evaluation and deployment. For the idea itself, read what agentic AI is.