Copilot Studio: what is new in multi-agent systems, Prompt Builder, and enterprise AI governance

Copilot Studio is entering a new phase: instead of focusing only on isolated agents, the platform is now betting big on connected multi-agent systems, with more native integrations, more control in production, and a much more mature building flow. Microsoft is rolling out orchestration capabilities, new Prompt Builder features, and governance improvements so companies can move from isolated pilots to AI scenarios that are truly integrated with the business.

In short, Copilot Studio helps organizations build and manage multiple agents that cooperate with each other, connected to existing data, apps, and processes. With that, it becomes possible to go beyond one-off chatbots and design an agent ecosystem that operates at scale, adapts to new requirements, and delivers real impact on productivity and business outcomes.

The main updates include:

• Multi-agent features in general availability, integrated with Microsoft Fabric, Microsoft 365 Agents SDK, and open Agent-to-Agent (A2A) protocols.
• An immersive Prompt Builder, available directly in each agent’s tools tab, to speed up experimentation and fine-tuning.
• More flexible governance and content moderation controls, designed for regulated scenarios and production usage.
• Enhancements in connectors, automated evaluation, Teams meeting agents, MCP integration, and new language models in paid preview.

Agents working together in the same ecosystem

Scaling AI inside a company goes way beyond building a single agent that works well. The real challenge is making multiple agents, created by different teams, collaborate reliably, reusing what already exists instead of reinventing everything for each new project.

A common scenario looks like this: the data team builds an analytics-focused agent, the apps team builds another agent for internal workflows, and the productivity team creates a third one for day-to-day tools. On their own, each agent adds value, but when a flow requires:

• knowledge from one system,
• reasoning from another,
• and action in a third,

integration becomes a festival of workarounds, custom APIs, and fragile handoffs. That slows down adoption, makes the jump from pilot to production harder, and drives up maintenance costs.

To address this, Copilot Studio is making several multi-agent orchestration capabilities generally available, with the goal of turning a set of agents into a coordinated system, not just a collection of islands. Highlights include:

• Integration with Microsoft Fabric to work with data and analytics at scale.
• Orchestration with Microsoft 365 Agents SDK to combine agents designed for Microsoft 365 experiences.
• Agent-to-Agent (A2A) communication via an open protocol, enabling interaction with agents from other platforms.

Multi-agent with Microsoft Fabric: data and analytics at the core

With multi-agent support for Microsoft Fabric, agents built in Copilot Studio can collaborate with Fabric agents to reason over corporate data at scale. Instead of treating every data-intensive use case as a new engineering project from scratch, business agents can connect directly to the company’s data estate.

In practice, that means:

• Agents for support, service, or sales can query and analyze data in lakehouses, warehouses, and other Fabric resources.
• Business context is no longer limited or detached and starts to reflect the organization’s real-world reality much better.
• Responses become more accurate, relevant, and actionable, because the agent understands the full scenario, not just a slice of it.

Instead of burning time on one-off integrations, teams can compose experiences using building blocks already available in Fabric and Copilot Studio, keeping the focus on business logic.

Multi-agent with Microsoft 365 Agents SDK: reuse instead of rebuild

Another key pillar is multi-agent support with the Microsoft 365 Agents SDK. With it, teams can orchestrate Copilot Studio agents alongside agents built for Microsoft 365 experiences.

Instead of duplicating the same logic everywhere (fetching data, applying internal rules, closing repetitive tasks), teams can reuse capabilities they have already built and combine them into new flows. That helps to:

• Reduce duplicated effort, since a single agent can be reused in more than one context.
• Create flows across apps, such as an agent that operates in Teams but also triggers logic already built for Outlook or SharePoint.
• Keep the user experience more consistent and efficient at the front end.

The result is an ecosystem where agents become pieces of a corporate LEGO set: each one does something well-defined and can plug into different flows without rewriting everything.

Agent-to-Agent (A2A): open protocol for agents from any stack

With Agent-to-Agent (A2A) support, Copilot Studio agents can communicate directly with other agents, whether from Microsoft, partners, or third parties, using an open protocol.

This is crucial: the future of enterprise AI is not going to revolve around a single technology stack. Companies will use different tools, vendors, and agents built on varied platforms. That is why Copilot Studio needs to participate in a broader ecosystem, not get locked into a single product.

A2A enables exactly that:

• An agent in Copilot Studio can delegate tasks to an external agent and receive the result back.
• Teams can combine the best of breed from different platforms without locking everything into a single stack.
• Interoperability becomes part of the solution design, not an after-the-fact patch.

Real-world impact: the Ask Microsoft case and other scenarios

A concrete example of this architecture is Ask Microsoft, the web agent Microsoft uses as both showcase and internal lab. In the beginning, it ran on a single-agent architecture. As traffic grew and knowledge sources multiplied, that setup started to get overloaded, pushing up response times.

With Copilot Studio, the solution migrated to an approach with generative orchestration and multiple specialized sub-agents. Today, the flow works like this:

• A primary agent receives the user’s question.
• Specialized sub-agents handle areas like Azure, Microsoft 365, pricing, trials, and more.
• The primary agent orchestrates the responses, ensuring a single, coherent, and fast outcome.

This design lets Ask Microsoft answer complex questions spanning multiple products and still adapt answers based on the section of the site where the user is browsing. The perceived impact is a smoother experience that is ready to scale without abrupt performance degradation.

Another classic scenario is a bank: the loan team has an agent that handles mortgages, while the accounts team uses another one for balance and transaction questions. For the customer, though, the expectation is obvious: they want a single experience without having to talk to two different systems.

With multi-agent orchestration, each agent remains a specialist in its own domain, but the user interacts with just one interface. If someone asks about their mortgage installment and checking account balance in the same conversation, the system combines answers from both agents and delivers a coherent response without forcing the customer to switch channels or repeat information.

The same logic applies in other industries. One example Microsoft highlights is Coca‑Cola Beverages Africa, which already uses Copilot Studio agents together with Microsoft Dynamics 365 to run planning cycles and automate end-to-end workflows. The reported benefit is about 1 to 1.5 hours saved per planner, every day, thanks to coordinated automation.

According to Microsoft, these multi-agent capabilities will be fully rolled out to all eligible customers by April 2026, cementing this model as the platform’s standard.

Faster, more controlled prompt authoring

As agents grow more sophisticated, the prompt that defines each agent’s behavior becomes a critical component. A good prompt leads to useful answers; a great prompt unlocks the agent’s full potential. The problem is that, until recently, the prompt tuning cycle was pretty clunky.

Previously, whoever was building the agent had to leave the main flow, open another editor, make a small change, test, go back, tweak again, and so on. In environments where the agent serves specific business scenarios, this back-and-forth created friction and slowed down evolution.

To fix this, Copilot Studio introduced the immersive Prompt Builder, now generally available. The idea is simple: move prompt editing directly into each agent’s tools tab.

Within this experience, you can:

• Update instructions and the operational context for the agent.
• Switch the model used by the prompt.
• Add inputs, parameters, and knowledge sources.
• Test changes in real time without leaving the agent screen.

This helps keep context during tuning. Instead of getting lost in multiple windows, the person configuring the agent can iterate quickly, with a clear view of how that prompt fits into the bigger picture.

Tools we use daily

Translation

Text Inspection & Clipping

Social Media

Search & Research

Recruitment

Programming

Productivity & Organization

Operations

Marketing & Sales

Finance & Investment

Education

Design & Media

Data & Analytics

Content & Writing

Automation

Example: prompts in sensitive domains like healthcare

The benefits are even clearer in areas dealing with sensitive domains or highly specific terms, such as healthcare. Imagine an agent designed to support clinical documentation:

• It must handle medical terminology correctly.
• It has to comply with strict privacy and security policies.
• It cannot be over-blocked by generic content filters.

In Prompt Builder, the team can adjust instructions, swap the knowledge source for a more suitable one, calibrate the answer style, and test outputs with terms that often trigger default safeguards. All of that happens in a single workspace, reducing friction and effort to reach a configuration that is production-ready.

More prompt options: content moderation and model selection

Speaking of filters, Copilot Studio also added content moderation settings specifically for prompts, now generally available in supported regions. This provides more control over how sensitive the managed models should be to potentially harmful content.

In some sectors, such as healthcare, insurance, and public safety, default settings can be too strict for the type of information being processed. With the new controls, teams can fine-tune this sensitivity and unlock legitimate scenarios without losing the baseline layer of safety.

On top of that, the Prompt Tool now supports Anthropic Claude Opus 4.6 and Claude Sonnet 4.5 in paid preview in the United States. The goal is to expand choice so teams can better match each prompt with the most suitable model, instead of forcing everything into a single trade-off profile between performance, reasoning depth, and cost.

Taken together, these improvements speed up the prompt experimentation cycle while keeping the flexibility and control required for production scenarios.

Other Copilot Studio updates

Beyond multi-agents and Prompt Builder, Microsoft also shipped a range of updates across connectors, evaluation automation, meeting capabilities, and model support.

• Improved ServiceNow and Azure DevOps connectors: now generally available, with better understanding of operational questions, stronger ticket and work item retrieval, and more complete and actionable responses.
• Evaluation automation APIs: now available via Microsoft Power Platform APIs and connectors, enabling programmatic evaluations and embedding quality checks into CI/CD pipelines.
• Meeting agents in Microsoft Teams: agents can access real-time transcripts and group chat, supporting scenarios like answering questions during meetings, surfacing relevant information at the right moment, and helping capture decisions and follow-ups.
• Support for apps via Model Context Protocol (MCP) and Apps SDK: expands ways to connect agents to external work apps, making it easier not only to retrieve information but also to execute actions in business systems.
• New supported models: including Grok 4.1 Fast, GPT-5.3 Thinking, and GPT-5.4 Instant in paid preview, offering more options to balance speed, cost, and capability.

Overall, Copilot Studio’s direction is clear: move beyond isolated AI experiences and push toward connected, governed, and integrated systems that align with how work already happens inside companies. As multi-agent ecosystems grow, these features help teams reach more channels, cover more knowledge sources, and respond more precisely to business needs.

Looking ahead, Microsoft is already signaling further updates through April 2026, including more capabilities in voice, workflows, and building experience, reinforcing the path of turning Copilot Studio into a central orchestration layer for agents across the organization.