An assistant writes a class that talks to Cosmos DB in seconds. Whether it fits the access patterns, the key and the sign-in of the Kladde, it knows only if you tell it, and whether it is correct, in the end only the person who checks it knows. For this, Learn lists the Azure Cosmos DB Agent Kit and names the capability "Implement Agent Kit to follow Azure Cosmos DB development recommended solutions". This lesson summarizes what Learn says about it and builds a checklist for code from it, no matter whether a person wrote it or a model.
Not run against Azure, the kit not installed: the statements about the kit come from the Learn pages alone; whether the kit's rules in the repository are as good as Learn describes them was not checked (open). The C# code is compiled with Microsoft.Azure.Cosmos 3.64.0 and was run locally. This lesson changes none of the facts F-01 to F-08.
What the kit is
The kit is a collection of agent skills, that is, instructions that an assistant loads when needed, for data model, partition key, queries and SDK use. Learn names as topics, among others, key choice, hot partitions, queries, indexing, throughput, distribution, security, monitoring, vector and full-text search. The source is a repository on GitHub (AzureCosmosDB/cosmosdb-agent-kit), the path of choice is npx skills add AzureCosmosDB/cosmosdb-agent-kit with Node.js; for individual hosts, Learn names paths of their own (Claude Code, GitHub Copilot, Gemini CLI, Codex, Cursor, more). After that, you check whether the host finds the capability cosmosdb-best-practices; an answer that mentions "Cosmos DB" does not prove that. The installation does not need an Azure subscription.
Learn draws three limits that you should keep in mind. The kit gives advice; it does not sign the assistant in to Azure and gives it no access to a database. The Microsoft Learn MCP server is something else: it searches documentation and delivers no data. Database access needs specially set up tools with their own permissions. And advice is not a measurement: check generated code, test in your environment and measure before you claim an improvement.