Microsoft has just taught his AI agents to talk to each other – and he could transform our way of working

Microsoft has announced a significant expansion of its Copilot Studio platform at Build 2025 today, introducing multi-agent systems that allow different AI agents to collaborate on complex commercial tasks, as well as new developer tools, safety improvements and integration with WhatsApp.

The sequence of announcements represents Microsoft’s most ambitious attempt to make AI agents more practical for the use of the company, approaching key limitations that have hampered the wider adoption of agents technology in commercial areas.

“We see customers carrying out large-scale production deployments that governance and observability become even more critical,” said Ray Smith, vice-president of AI agents at Microsoft, in an exclusive interview with Venturebeat. “The beauty of Copilot Studio is that we provide a managed infrastructure framework with integrated life cycle management and complete governance capacities.”

How the new multi-agent system from Microsoft transforms corporate workflows

At the heart of the announcements is the new multi-agent system of Microsoft, which allows agents built with Copilot Studio, Microsoft 365, Azure AI Service and Azure Agents Manufacture of working together, by delegating tasks to each other to finish complex commercial processes.

This ability is a fundamental challenge that organizations were faced when implementing agents’ technology. “The creation of a reliable process in a single agent is extremely difficult,” said Smith. “Decomposing it into several agents not only improves maintainability and simplifies the strengthening of solutions, but it also considerably improves global reliability.”

The system allows scenarios such as a Copilot Studio agent drawing sales data from a CRM, putting it back to a Microsoft 365 agent to write a proposal in Word, then trigger another agent to plan follow -up in Outlook.

Microsoft also emphasizes interoperability thanks to the management of agent agent protocol recently announced by Google, potentially allowing the communication of multiplatform agent.

The “computer use” functionality provides a human -type user interface interaction to AI IA agents without dependencies

Another key ad is “the use of the computer” for Copilot Studio agents, which allows agents to interact with desktop applications and websites by controlling the interfaces directly – by clicking on the buttons, browsing the menus and tapping fields – even when the APIs are not available.

“When APIs are not available, this feature can interact directly with user interfaces – whether desktop applications, browsers or other platforms,” ​​said Smith. “It provides what we call” extensibility without clips “and operates according to the coordinates of intention rather than pixels, unlike traditional office automation. This approach focused on the goal makes it much more robust. ”

This capacity is currently available via the Microsoft Frontier program for eligible customers with more than 500,000 Copilot Studio messages and an American environment.

Customizable model selection and Python analysis Solutions AI SuperCharge Ai

Microsoft offers organizations more flexibility with their AI models by allowing them to provide personalized models from the Azure AI foundry in Copilot Studio. This includes access to more than 1,900 models, including the last of Openai GPT-4.1, Llama and Deepseek.

“Start with standard models because they are already fantastic and improve continuously,” said Smith. “Companies generally choose to refine these models when they have to integrate a specific domain language, unique use cases, historical data or customer requirements. This personalization ultimately stimulates greater efficiency or better precision. ”

The company also adds an code interpreter functionality which provides python capacities to studio agents Copilot, allowing data analysis, visualization and complex calculations without leaving the Copilot Studio environment.

Smith highlighted financial applications as a particular force: “In financial analysis and services, we have found a remarkable breakthrough in the last six months,” said Smith. “The models of deep reasoning, fueled by the learning of strengthening, can actually self-make any process that produces quantifiable outings.”

He added that these capacities excel in “a complex financial analysis where users must generate code to create graphics, produce specific results or perform detailed financial assessments.”

Whatsapp integration extends the scope of the AI ​​agent to billions of world users

From early July, organizations will be able to publish Copilot Studio agents in Whatsapp, allowing them to reach customers via one of the most popular messaging platforms in the world.

“WhatsApp is obviously a key chain. Globally, it’s quite huge,” said Smith. “This has therefore become a high priority for us through the different channels and integrations, to unlock this as a means for end users to interact at a time that suits them.”

This addition, as well as a new SharePoint canal (now generally available), considerably widens the scope of personalized agents beyond the Microsoft ecosystem.

Microsoft Bridges The developer’s experience gap with VS code extension and improved administrative controls

For professional developers, Microsoft is launching an extension of Visual Studio code for Copilot Studio, providing familiar tools and workflows to the development of agents. The extension provides features such as Intelisense, formatting colors and “finding all references”, allowing developers to modify agents directly from the Visual Studio code.

IT administrators also obtain new tools, including a centralized “Agents & Connectors” page in the Microsoft 365 Admin Center for full life cycle of co -pilot agents. This interface allows administrators to visualize all agents, filter by metadata, allocate sensitivity labels, manage the behavior of the connector and block or remove agents that violate security policies.

The discovery of agent fueled by AI resolves the “What agent do I need?” Problem for end users

While organizations are developing more specialized agents, Microsoft takes up the challenge of discovery of agents with new recommendations from conversation agents. This feature suggests the most relevant agent based on the needs of a user, dynamically recommending transfers if necessary.

“It is fundamentally as a model of agency cloth,” said Smith, referring to the generation with recovery. “He uses vector databases and indexing according to the agent’s description. When you describe the task, it’s research and goes:” Hey, I think you should watch this agent. “”

The system recommends only to agents to which the user has access, respecting the existing authorization structures.

Overcoming the obstacles of the implementation of the company’s AI: Microsoft’s roadmap for success

Despite the new capacities, organizations are still faced with major challenges when implementing multi-agent systems. Smith recommends focus on specific use cases rather than trying full transformations from the start.

“Best practices have been clients who focused on use cases or a problem to start, broken down into games and have sought to integrate an agent into various parts of this process,” he said. “They go from an agent to two agents, then they put it together in a multi-agent solution.”

He highlighted the importance of the practical experience with the tools: “You must know well in what it is not … It is a mixture of tools that you will use to transform your commercial process.”

Microsoft’s business AI strategy positions him to direct the next wave of transformation into a workplace

These announcements collectively strengthen Microsoft’s position on the company’s AI market, where competition with Google, Amazon and specialized AI startups continues to intensify.

By meeting the main requirements of the company such as security, governance and interoperability, while simultaneously expanding the capacities of the platform thanks to functionalities such as the use of computers and the interpretation of the code, Microsoft creates a more complete offer for organizations that seek to deploy large-scale AI agents.

“Microsoft is largely in a leading position in the construction and management and scaling of these agency or agent workflows,” said Smith. “I think we have reached an important inflection point for some of the main features of customers.”

The double approach of the company – by taking care of both code -free tools for professional users and professional development environments for engineers – also distinguishes it from competitors who mainly focus on developers or professional users.

“Our platform serves experienced developers who prefer to use Visual Studio for direct access to APIs and executives,” said Smith. “It also welcomes professional users with a limited coding experience that wishes to start with a specific commercial problem and prototype solutions quickly using natural language.”

While Microsoft has taken up many key challenges for the adoption of corporate AI with these announcements, Smith suggests future developments by focusing on the discovery and accessibility of agents.

“I think the next part will be all these agents surface in a market or catalog, then accessible from AI assistants like Copilot or Microsoft 365 Copilot and beyond,” he said.

The Build 2025 announcements represent a significant step forward for Microsoft’s strategy of Microsoft, but the company continues to face the challenge of stimulating a generalized adoption of these technologies in the industries and helping organizations to carry out a tangible commercial value from their IA investments.

As Smith concluded, “at the heart of any transformation is not really technology itself. It is a question of providing solutions which are more profitable, faster, of better quality and which ultimately create a superior customer experience. ”