Learn to build autonomous AI agents

Python, Git, Linux, CLI, networking, HTTP, REST APIs, Docker, SQL, testing, and CI/CD — the engineering base every AI builder needs.

What GenAI is, how LLMs work, tokens, embeddings, prompt engineering, and core concepts — without repeating RAG or tool-calling deep dives.

Neural networks, attention, encoders/decoders, BERT, GPT, KV cache, RoPE, MoE, and quantization — interview-depth ML intuition. Skip if you are focused on building agents, not training models.

Hands-on with OpenAI, Claude, Gemini, Ollama, open-source models, streaming, and multimodal APIs.

Document loaders, chunking, vector DBs, hybrid search, re-ranking, LangChain, ChromaDB, and Streamlit demos.

What agents are, how they work, planning, reasoning, reflection, and building your first agent without frameworks.

Working, short-term, long-term, semantic, and episodic memory — one of the biggest interview topics for production agents.

How an agent builds the next model call: assemble, select, compress, isolate, route, and budget context from instructions, memory, retrieval, tools, and state — not just a static prompt.

Function calling, JSON mode, structured outputs, tool registries, permissions, and building a tool-using assistant.

MCP as the agent ↔ tools/data protocol: host/client/server, tools/resources/prompts, sampling/roots/elicitation, local stdio vs streamable HTTP, authorization, and how to securely consume third-party servers.

Why frameworks exist, how to pick one, and awareness of the rest of the landscape. AutoGen and Semantic Kernel are legacy/migration. Deep dives follow as their own phases. Protocol stack: MCP ↔ tools, A2A ↔ agents, AG-UI ↔ users.

Visual-first graph runtime: nodes, state, routing, tools, checkpoints, HITL, streaming, time travel, subgraphs, Platform — then one support-agent project that uses all of them.

OpenAI's official agent runtime: agents, tools, sessions, handoffs, guardrails, tracing, and a shipped assistant.

Anthropic's production agent SDK: the same loop and context-management ideas behind Claude Code — tools, permissions, hooks, MCP, subagents, sessions, sandboxing, and a coding agent.

Crews + Flows: why CrewAI exists, the agent loop under the roles, tasks, tools, processes, delegation, memory, knowledge, and production orchestration — then one beginner Research Crew you run locally.

Type-safe Python agents: model-agnostic runtime, typed tools, dependencies, structured results, and a small production agent.

Microsoft Agent Framework is the successor to AutoGen and Semantic Kernel. Learn agents, workflows, durable execution, HITL, MCP, and A2A. AutoGen remains as legacy/migration awareness (same module slugs, so existing progress is kept).

Google's Agent Development Kit 2.x: agents, graph-based workflows, dynamic delegation, A2A, sub-agents, and a Gemini-backed agent you can run locally.

ReAct, Plan & Execute, Reflexion, Tree of Thoughts — learned after building agents, not before.

A2A as agent ↔ agent interoperability, supervisor/worker patterns, coordination, and trust between agents.

How to prove an agent works: datasets, LLM-as-judge, online/offline eval, benchmarks, agent-specific OpenTelemetry, and regression suites.

OWASP 2026 agentic threats: goal hijacking, tool misuse, identity/privilege abuse, supply chain, memory poisoning, sandbox escape — plus identity, least privilege, and governance.

The runtime that makes an agent operable: durable execution, sandboxing, state, cancellation, model routing, versioning, and the usual serving stack.

Agent ↔ user protocol: streaming events, shared state, tool-call UI, HITL approvals, generative UI, and frontend integration. MCP is tools, A2A is agents, AG-UI is the human.

Playwright, browser automation, computer use, form filling, and web navigation agents.

STT, TTS, realtime voice, and agents that understand images, audio, video, PDFs, and screens.

Enterprise RAG, knowledge bases, RBAC, compliance, identity, audit logs, data lineage, and human approval at scale.

Repository understanding, sandboxed shell, patches, tests, git/PR workflows, AGENTS.md, and coding-agent evaluation.

Enough model engineering to operate agents: fine-tuning awareness, LoRA/QLoRA, distillation, inference optimization, serving. Image/video/research stay for completeness — this is not an ML-researcher track.

Production-ready portfolio projects — AI software engineer, research assistant, customer support, and more.

Agent system design, LangGraph, MCP, multi-agent, memory, context, runtime, AG-UI, production debugging, and mock interviews.