// the find
NirDiamant/agents-towards-production
End-to-end, code-first tutorials for building production-grade GenAI agents. From prototype to enterprise deployment.
A collection of 28 Jupyter notebooks teaching you how to ship AI agents to production — covering deployment, memory, observability, security, and multi-agent coordination. It's aimed at Python developers who have a working agent prototype and want to understand the operational layer: containerization, GPU scaling, guardrails, tracing. Each tutorial is self-contained and sponsored by a vendor whose product it features.
The sponsor-funded tutorial model means each notebook is written with actual product knowledge — the Redis memory tutorial, for instance, walks through dual short-term/long-term architecture with real vector search, not a toy example. Coverage of the full production surface is genuinely broad: LlamaFirewall for prompt injection defense, LangSmith for tracing, IntellAgent for behavioral evaluation, and A2A protocol for multi-agent communication all in one place. The Kotlin/Koog tutorial is a rare find — almost everything in this space is Python-only, and having a JVM-native path matters if you're building on a Java/Kotlin backend. The file-conversion tutorial includes an honest fidelity benchmark against competing extractors, which is the kind of practical callout most tutorial repos skip.
Almost every tutorial is a sponsored placement, which means the tool selection is driven by who paid rather than what's best — you won't find any critical comparison of LangSmith vs. alternatives or an honest assessment of when you should not use Redis for agent memory. The custom non-commercial license is a real constraint: you can learn from these notebooks but can't use them as a basis for a product, which undercuts the 'production-grade' framing. There's no cohesive codebase tying the tutorials together — each one is a standalone notebook with its own dependencies, so there's nothing showing how these pieces actually compose in a real system. The 'production' label is also aspirational in places; the Docker tutorial is a basic intro to containers, not anything specific to agent workload characteristics like long-running processes or state persistence across restarts.