Agents Towards Production

What it covers

The tutorials are grouped by the production concern they address:

• Tool integration and protocols: secure tool calling, the Model Context Protocol, and multi-agent communication with the A2A protocol.
• Memory: dual-memory designs with semantic search, self-improving memory and knowledge-graph memory.
• RAG and data: durable ingestion pipelines, production RAG agents, web data collection and real-time search, and local file conversion.
• Deployment: containerising agents with Docker, deploying agents as APIs with FastAPI, running models on-premises with Ollama, managed deployment on AWS, and GPU hosting.
• Security: agent guardrails and hands-on security evaluation.
• Observability and evaluation: tracing and debugging, and automated behavioural evaluation.
• Frameworks and UI: stateful workflows with LangGraph, agents in Kotlin, fine-tuning agents for a domain, and a simple Streamlit chat front end.

The repository had about 21,500 GitHub stars as of September 2026 and was updated that month.

Understand the sponsorship model

Many tutorials are produced with, and centred on, a specific company's product, and the repository lists these as tutorial sponsors. That is how it stays free, and the tutorials are practical, but it means each one shows one vendor's way of solving a problem. Treat them as worked examples of a concern (memory, tracing, guardrails), then compare with open source or other options before committing your stack.

The repository uses a custom non-commercial licence, so read it before reusing code in a commercial product.

Why this layer matters

Most agent material stops when the demo works. Production problems are different: the agent forgets context between sessions, a tool call leaks data, costs spike, a failure cannot be reproduced, or nobody can tell whether a prompt change made things better. This collection is one of the few free resources organised around exactly those problems.

How to use it

Do not work through it from top to bottom. Instead:

1. List what your prototype is missing: persistent memory, a deployable API, tracing, security checks or evaluation.
2. Pick the one tutorial for your biggest gap and follow it with your own agent rather than the example.
3. Add tracing early; it makes every later step easier to debug.
4. Do a security tutorial before giving any agent tools that change real systems.

Before you start

You should already be able to build a basic agent with tools in Python, for example after Microsoft's AI Agents for Beginners or the Hugging Face Agents Course. Some tutorials need accounts with the featured services, which usually have free tiers; check each before starting.

Common questions

Is it a course? No. It is a set of independent tutorials; there is no fixed order or certificate.

Can I use the code at work? Check the licence first, since it is non-commercial by default.

Best for: developers turning a prototype agent into something dependable enough for real users.