Microsoft unveils Foundry overhaul for managing, optimizing AI agents

Microsoft has announced a slew of updates to Microsoft Foundry to simplify the creation and management of AI agents throughout enterprise environments, unifying observability, security, and developer pipelines.

Chief among the new announcements are new connectors and data unification layers to help enterprise customers improve the functionality of their AI tools.

Foundry IQ and Fabric IQ will help IT administrators and developers to ground AI agents in their enterprise data and keep track of deployed systems.

The former is a new knowledge system intended to connect agents to any data in OneLake, with interconnectors such as Amazon S3 and Snowflake.

ITPro spoke to Eric Boyd, CVP AI Platform at Microsoft, who explained that Foundry IQ is the next-generation of retrieval augmented generation (RAG), empowering AI agents with relevant, efficiently-sourced information for accurate outputs.

“Foundry IQ is looking to give you a single tool that will help you manage across all these different data sets that you have,” he said.

Fabric IQ brings together deployed systems within a comprehensible, semantic cloud showing how they relate to other systems within an enterprise environment.

Boyd explained that this helps the problem faced by many IT organizations beginning to manage thousands of agents, who are now forced to tackle greater levels of complexity, reduced oversight, and concerns about rising costs.

“We would talk to a company a couple years ago and they would say, ‘I'm struggling to get my first AI application into production’,” he told ITPro.

“Now they say ‘I have a thousand of them, I don't know who created this, I don't know how to manage it, I don't know how much I'm paying for it’. And so [they’re] just looking for some help with all of that.”

The new announcements in Microsoft Foundry, including Fabric IQ and Foundry IQ, are the hyperscaler’s answer to this, acting as a unified platform through which developers can create agents and AI applications, as well as manage them.

Foundry Control Plane, another major announcement, extends the capabilities of Agent 365 to give developers direct control over agent lifecycle management, behavioral guardrails, observability, and policies.

Rooted in Microsoft Defender and Entra Agent ID, the platform will help IT teams to enforce safety and security controls across their fleet of AI agents, including runtime threat detection, ownership and access controls, and usage limits.

“Every agent that gets created using Microsoft tools will be visible there, agents that are created outside of Microsoft tools with a couple of lines of code can be added, and so then the telemetry will flow through, and you can start to see those as well,” Boyd said.

“So that visibility across everything that's happening at the enterprise gives the enterprise IT leader the ability to understand what's happening – and then to work to control it.”

Foundry Control Plane is now available in preview. In the near future, Microsoft will add an integration between the service and Palo Alto Networks.

Microsoft is also previewing a unified catalog of model context protocol (MCP) tools, including logic app connectors for third-party tools such as HubSpot, Salesforce, and SAP as well as off-the-shelf AI services for transcription and documents processing.

In addition, developers will be able to use the catalog to search and manage private MCP tools from a unified dashboard within Microsoft Foundry.

The right model for the job

A new model router function within Microsoft Foundry is now generally available, to automate the process of choosing the right AI model for the right task, with a focus on balancing results with inference costs.

Boyd told ITPro that the new feature is especially important because enterprise customers can struggle to pick the right model for their applications.

While some business problems can easily be resolved by smaller models, at a relatively low cost, others require the most sophisticated models – including reasoning models – which come with a higher associated cost.

The model router helps developers to manage which to use for which tasks by dynamically switching between models for each user prompt.

Models are chosen based on the complexity of this prompt versus their potential cost and latency.

According to Microsoft data collected from customers with early access, the model router achieves a 50% reduction in model costs and a 40% improvement in response times.

“When we first started with this journey a few years ago, there were just like two or three models that mattered,” Boyd said.

“Now we have 11,000 models in the Foundry Model Catalog, so it's beyond what a single developer can figure out, and figure out how they can manage and choose across all of them.”

At time of launch, the model router can choose from the following models:

• DeepSeek-v3.1
• GPT-4.1 family
• GPT-5 family
• GPT-oss-120b
• Grok-4
• Grok-4-fast
• Llama-4-Maverick-17B-128E-Instruct-FP8
• Llama-33-70B-Instruct