Agentic AI

Understand the defining characteristics that separate an AI agent from a simple chatbot or API call — autonomy, goal-directedness, environment interaction, and feedback loops.

Learn how large language models generate text, what the context window is, how temperature and sampling affect output, and why these properties matter when building agents.

Master structured prompting techniques — system prompts, few-shot examples, chain-of-thought, and output formatting — that form the backbone of agent instruction.

Survey the landscape of agent designs: reactive agents, deliberative agents, BDI (Belief-Desire-Intention), and modern LLM-based agent architectures.

Understand how breaking problems into explicit reasoning steps improves agent accuracy, and learn techniques like CoT, self-consistency, and tree-of-thought.

Learn how agents break complex goals into subtasks, create execution plans, and adapt plans when intermediate steps fail or produce unexpected results.

Explore how agents can evaluate their own outputs, detect errors, and iteratively refine their work through self-reflection loops like Reflexion and self-debugging.

Learn the Reasoning + Acting loop where agents interleave thinking steps with tool calls, observe results, and decide the next action — the dominant pattern in modern agents.

Understand how agents invoke external functions — API design for tools, JSON schema definitions, parameter extraction, and handling tool responses.

Learn how agents write and run code to solve problems, including sandboxed execution environments, error handling, and iterative debugging workflows.

Explore how agents navigate websites, call REST APIs, and extract structured information from unstructured web content to accomplish real-world tasks.

Learn how agents use vector databases and retrieval systems to access large knowledge bases, ground their responses in facts, and reduce hallucination.

Learn strategies for managing limited context windows — summarization, sliding windows, priority-based truncation, and when to offload information to external storage.

Understand how agents maintain conversation state, scratch pads, and intermediate results within a single task execution session.

Explore techniques for agents to remember information across sessions — vector stores, structured databases, memory retrieval strategies, and memory consolidation.

Learn common patterns for multiple agents working together — supervisor/worker, debate, ensemble, pipeline, and blackboard architectures.

Understand how agents exchange information — message passing protocols, shared state, structured handoffs, and maintaining coherence across agent boundaries.

Explore how to design agents with focused expertise and route tasks to the right specialist — creating systems that are more capable than any single agent.

Learn the LangChain ecosystem for building agent chains and LangGraph for creating stateful, graph-based agent workflows with cycles and branching.

Build agents using Claude's native tool use, the Anthropic SDK, and the Claude Agent SDK — including structured outputs, streaming, and extended thinking.

Explore OpenAI's approach to agent building with their Agents SDK, Assistants API, function calling, and code interpreter tools.

Learn to build agents from scratch without frameworks — understanding the core loop of prompt → LLM call → parse → tool execution → observe → repeat.

Learn how to constrain agent behavior — input/output validation, action allowlists, human-in-the-loop approvals, and preventing prompt injection attacks.

Understand how to measure agent performance — task completion rates, tool use accuracy, reasoning quality, and established benchmarks like SWE-bench and GAIA.

Learn to trace agent execution, log reasoning chains, monitor tool calls, and debug failures in complex multi-step agent workflows.

Explore strategies for managing token usage, caching, model routing (using smaller models for simple steps), and reducing end-to-end agent latency.

Design systems where agents escalate to humans for approval, clarification, or quality review — balancing autonomy with oversight for high-stakes tasks.