Tutorials
A guided path from a one-line chat to served, packaged, connected agentic systems. Lessons 1 to 8 teach the native runtime; 9 to 11 take a system outward.
| # | Lesson | What you will learn |
|---|---|---|
| 1 | Quick Start | Tactic.quick(), open → receive → respond, local models, lllm create |
| 2 | Agents and Dialogs | The core mental model: Agent, Dialog, multi-turn chats, forking |
| 3 | Prompts and Structured Output | Templates, parsers, Pydantic output, exception handlers, prompt inheritance |
| 4 | Tools, Proxies, and Sandbox | @tool, the tool loop, MCP, proxies with run_python and CALL_API |
| 5 | Native Tactics | Multi-agent programs, call sessions, sub-tactics, batch and async |
| 6 | Packages and the Hub | lllm.toml, discovery, YAML agent configs, validating and publishing |
| 7 | Logging and Cost Tracking | Loggers, call sessions and cost, observers |
| 8 | Advanced Patterns | Proxy orchestration, forking, pipelines, streaming, offline testing |
| 9 | Serving and Composition | lllm serve, the HTTP API, remote tactics, refs, lllm.Resolver |
| 10 | Pydantic AI Agents | Pydantic AI agents as tactics, tactics as Pydantic AI tools |
| 11 | Channels with SSSN | Channels, TacticProcessor, serving a store, AAAX strategies |
Prerequisites
- Python 3.10+
pip install "lllm-core[native,server]"- A model: an API key for a hosted provider (
OPENAI_API_KEY,ANTHROPIC_API_KEY, ...) or a local OpenAI-compatible server such as vLLM, SGLang, or NVIDIA Dynamo. Lesson 1 shows both.
How to read them
Each lesson builds on the previous one, and every code block is runnable. Lessons that call a model say so; everything else runs offline. When you want the full detail behind a lesson, each one links to its reference page.