Case study · Nebius Agent Blueprint Recipes
Agents ready to outgrow the demo.
Ten open-source recipes, one reference blueprint and a catalogue that demonstrate—in code—what production-ready AI agents actually require.
nebius-agent-blueprint-recipes / catalogue- Client
- Nebius
- Scope
- 10 recipes · 1 blueprint · catalogue website
- Timeline
- One month · press presentation on June 9, 2026
- Clément
- Clément SauvageProduct Engineer & Technical Lead
- Role
- Technical architecture · Development · Reference implementation
- Technology
- LangGraph · FastAPI · MCP · Pinecone · Tavily · Nebius
The context
The API works. The system still needs to be built.
An agent must find the right information, remember, refuse unsafe work, trigger real actions and make its behavior measurable. For Nebius, those capabilities live across different partners: the platform’s value appears in how they work together.
Do not document a promise. Ship executable proof.
The brief
One month to build a reference that could stand up to developers.
June 9, 2026 · San Francisco
Developers and specialist press
MIT · open source
Stripe · Clever Cloud
The approach
Recipes, not demos.
Each step introduces the next problem a team meets when turning a prototype into an operable service.
Foundation
Streaming, FastAPI, logs and metrics from the first line.
Knowledge
Pinecone knowledge with typed citations.
Grounding
Fresh web data with Tavily.
Orchestration
A typed LangGraph with bounded work.
Thread Memory
Memory inside one conversation.
User Memory
Persistent preferences with Postgres.
Observability
Every run inspectable in LangSmith.
Guardrails
Unsafe behavior is stopped.
Actions
Stripe payment link through MCP.
Simulation
Scenarios and failures scored with Snowglobe.
Production-shaped
The foundation starts at recipe 01.
Every recipe is a standalone FastAPI application: typed, observable, containerized and tested.
Pydantic & type hints
Structured JSON logs
Prometheus & healthchecks
Security and graceful shutdown
Docker & offline tests
Le blueprint
Sentinel shows the finished system.
A recipe teaches one concept. Sentinel combines them in a compliance auditor that checks 200 internal procedures against 36 frameworks, then files identified gaps in Jira.
The real difficulty
Ten autonomous projects speaking with one voice.
CI validates, lints and tests every project independently. A regression in recipe 07 cannot break 03; an eleventh recipe gets the same guarantees in one command.
Results
On time. Adopted by the team. Maintained without dependency.
The project was presented to the press on June 9. Nebius engineers now maintain it as a community resource.