Agents

An agent is a model with tools and a loop: the LLM decides which tool to call, your code runs it, the result goes back to the model, and the cycle repeats until the model stops asking. Every framework difference is noise until you have built the loop once. The AISL material makes you build it — then shows what the frameworks, protocols, and durable-execution systems add.

The loop from first principles

The Agentic Loop from Scratch starts with a fixed RAG pipeline, adds function calling one round at a time, automates the loop, and refactors it into a small framework with iteration caps. Building a Coding Agent with Python and Django covers the same ground with the OpenAI Responses API before porting the agent to OpenAI Agents SDK, PydanticAI, Anthropic, and Z.AI.

Frameworks and MCP

Once the loop is familiar, frameworks remove the boilerplate: Building AI Agents with MCP, PydanticAI, and OpenAI compares toyaikit, OpenAI Agents SDK, and PydanticAI on the same FAQ agent, then moves the tools behind a FastMCP server so notebooks, Cursor, and VS Code can all reach them — the protocol gets its own page: Model Context Protocol. The AI Engineering Buildcamp module 3 does the same consolidation inside a curriculum, with PydanticAI as the production choice in later modules.

Durable execution

Short agents fail fast and cheap. Long ones need durability: A Deep Research Agent over YouTube Transcripts with Temporal turns transcript ingestion into a Temporal workflow with activities, workers, and retries, then runs the research agent itself — with a summarization subagent for long transcripts — inside a workflow that survives API failures.

Safety comes with the loop

A tool-using agent is also an agent that can be steered off topic. Input and output Guardrails have their own page, with two workshops dedicated to the mechanics.

Voice and managed platforms

Voice is the one modality the loop itself does not cover: Building a Voice Coach with ElevenLabs designs the coaching prompt first, provisions an ElevenLabs agent, and connects the voice conversation — then verifies the result instead of trusting the coding agent.

Where to go next

• Build the loop by hand: The Agentic Loop from Scratch
• Retrieval that feeds the loop: RAG
• Keep agents on topic: Guardrails
• Put the loop somewhere users can reach: Agent Deployment
• Coding agents — agents that write code — get their own page: Building Coding Agents