Tutorials | Crusoe AI Platform

📄️Tutorial: research agent with memory

Build an agent that computes with a real Python sandbox and remembers facts across separate conversations, from empty folder to clean teardown.

Load documents into a VectorDB index, then deploy an agent whose answers are grounded in those documents and cite the file they came from.

Publish a Python function as a hosted MCP server, call it directly with curl, then wire an agent to it and watch the model use the tool.

Wire a Pub/Sub topic to a function so every published message is delivered automatically, with retries and a dead-letter topic.

Create a machine identity, mint an API key, and deploy on every merge from CI — with the security rules spelled out.

Build a real-time weather assistant from published MCP tools and an agent that calls them.

A LangGraph agent that plans before it searches, remembers across sessions, calls MCP tools, and refreshes itself on a schedule.

Four specialists — researcher, writer, fact checker, editor — with per-agent MCP toolsets and one shared memory.

A bucket drop becomes searchable chunks — object-store trigger, Pub/Sub fan-out, chunking, embedding, VectorDB upsert.

Receive a webhook, suppress the sender's retries, and fan out to Pub/Sub — with the failure path that keeps a transient error from losing the event.

A cron-triggered Go function that checks a vector index against the truth, records the verdict, and logs in a shape you can query.

A CSV lands in a bucket, becomes clean rows on a topic, and the object is retired — with the two infinite loops that are easy to build by accident.

Two-stage retrieval through atomic MCP tools, fed by an object-store-triggered ingest — with the score that shows why reranking earns its call.