Agents vs project type vs capability host

[DennesTorres]

I understand there is an ongoing transition between AI Project (the projects which use AI Hub) and Foundry projects - the projects which only use foundry.

The recommendation is to focus on the Foundry projects, correct ?

All my tests to use custom azure functions, MCPs, etc, were on Foundry projects and were failing in AI projects. Is this correct, or should I push to use the same resources in AI projects as well?

The problem: The Foundry projects only allow the creation of agents if there is a capability host present. If there is no capability host, I can't create an agent - the option doesn't appear in the UI.

The capability host requires a cosmos db to work. Cosmos db is expensive - at least for simple tests, it consumes a lot. The AI projects consume much less.

I tried to use cosmos db serverless, but the foundry project doesn't allow it. The agent creation is disabled.

Is there a way to deploy a foundry project with support to agent creation but using less expensive resources than cosmos? A simple cosmosdb serverless could solve the problem.

[leestott]

Hi @DennesTorres Yes, the recommendation is to prioritize Foundry projects over legacy AI Hub/Projects. Foundry Projects are the future of Azure AI Foundry and offer better support for agentic workflows, evaluations, and modular orchestration. You can read more about Foundry Projects here: https://learn.microsoft.com/azure/ai-foundry/overview

Your experience is correct: custom Azure Functions, Model Customization Pipelines (MCPs), and other extensibility features are designed to work with Foundry Projects. These often fail or are unsupported in AI Projects, which use the older AI Hub architecture.

However, agent creation in Foundry Projects requires a capability host. This host depends on Cosmos DB for NoSQL not Cosmos DB serverless. Without a capability host, the agent creation option won’t appear in the UI. Details on capability hosts are available here:
https://learn.microsoft.com/azure/ai-foundry/how-to/configure-capability-host

Cosmos DB serverless is not supported for thread storage in Foundry Projects. The required containers (e.g., thread-message-store, agent-entity-store) need provisioned throughput. This is a known limitation and contributes to higher costs for testing. You can review Cosmos DB pricing here: https://azure.microsoft.com/pricing/details/cosmos-db/ Unfortunately there’s currently no supported workaround to enable agent creation in Foundry Projects without using provisioned Cosmos DB. AI Projects do allow agent creation without a capability host, but they lack support for many advanced features.

Benefits for Projects
If you're testing lightweight scenarios, AI Projects may be more cost-effective. For production or advanced workflows, Foundry Projects are recommended, even if they require Cosmos DB.

Things you might consider:

• Deploying Cosmos DB with minimal RU/s and scaling manually
• Setting up autoscale with alerts to monitor usage: https://learn.microsoft.com/azure/cosmos-db/provisioned-throughput-autoscale
• Requesting quota exceptions or discounts via Azure support if you're part of an educational or MVP program: https://learn.microsoft.com/azure/azure-portal/supportability/how-to-create-azure-support-request

[DennesTorres]

Thank you for the feedback.

These details are what I originally identified, but they seems very unpractical. The simple deployment was very difficult.

However, I was very surprised about a week ago when everything changed. Out of the blue the "Agents" option in the UI started appearing in simple projects, without the custom capability host.

I tested the creation, tested the link with MCPs, and it works! It seems to be a recent change.

But the foundry projects still have a lot of missing points which are very annoying. For example, any RAG implemention in the foundry projects is extremely difficult compared to what it was before.

In open ai service we could just link with an AI Search with the right parameters and the vector search would work. In foundry projects, it doesn't. This creates a terrible misconnection between Agents and existing AI solutions. For example, the "Chat with your data" accelerator provided by Microsoft.

I should be able to link an Agent directly to the AI Search used by a "Chat with your data" deployment and have the RAG working in the Agent - but it doesn't. However, it works in AI Hub projects.

This misconnection and missing features are annoying specially because there is low to no information about if they are being handled. For example, you replied to my message, but with information that changed a week ago without any announcement.