Today, we’re introducing closed-loop learning for agents connected to the Power Apps MCP server, starting with the data entry tool. Every correction a user makes through the Agent feed persists as structured memory. On future runs, the agent retrieves that memory and applies it. Over time, those corrections consolidate into organization-wide patterns the agent applies across tasks. The feedback loop runs automatically in production. Nothing to configure, no data pipelines to build.

How closed-loop learning works

Say a finance team has an agent to process vendor invoices. The agent uses the data entry tool in the Power Apps MCP server to extract supplier names, addresses, and totals from PDFs. Most fields come through correctly, and the agent correctly fills in “UK” from the document, but the organization requires it to be normalized to “United Kingdom.” A user corrects it in the Agent feed. After a handful of corrections like this, the pattern becomes the agent’s memory. The next invoice with “UK” gets “United Kingdom” automatically. Over time, the system learns the pattern: abbreviations like “USA” or “DE” get expanded too. The agent also learns by example: it retrieves similar memories to the current request and applies the same processing steps used previously.   

Think of it like a driving instructor in the passenger seat. A student merges without checking the blind spot; the instructor corrects them, and the student remembers it for next time, because they wrote it down. That’s the first layer: memory-based optimization, the mechanism that captures individual corrections and applies them on future runs.

Corrections get you through the driving test. Patterns make you a confident driver. That’s the second layer: Genetic-Pareto optimization, an evolutionary prompt optimization technique. Memory-based optimization generalizes from the corrections it remembers. Genetic-Pareto goes further, distilling those corrections into rules compiled into the agent’s instructions, so the principle becomes the agent’s default behavior rather than a pattern it must retrieve each time. Both layers are live today for the data entry tool.

Both mechanisms run with the rigor of a production machine learning pipeline. Unlike the memory features built into most AI assistants, which personalize the experience for an individual user, closed-loop learning improves task accuracy for everyone using your organization’s agent. The learning stays scoped to your business process and governed by your tenant.

Inside the research

The approach

Today’s techniques for adapting Large Language Models don’t close the feedback loop between deployment and improvement. Retrieval-augmented generation (RAG) surfaces documents at inference time but doesn’t learn from outcomes. Fine-tuning adapts the model but requires a training and deployment cycle for every update. Open-loop prompt optimization improves a fixed prompt offline but doesn’t incorporate live feedback.

Closed-loop learning closes the gap through two complementary mechanisms. Memory-based optimization captures user corrections as structured memories retrieved at inference time, giving the agent immediate recall. Genetic-Pareto optimization periodically distills those memories into generalized rules using evolutionary prompt optimization, so the agent applies what it’s learned to cases it hasn’t seen. The agent acts, a user corrects, the system learns, the next action improves.

Closed-loop learning builds on two open-source projects. Memory-based optimization takes inspiration from Memento, though our implementation has evolved significantly to fit the Power Apps MCP server’s architecture. Genetic-Pareto optimization is implemented via GEPA, integrated through DSPy, Stanford’s framework for programmatic LLM optimization.

The evaluation

The benchmark had to reflect enterprise data entry in practice. One of our early customers, the UK Electoral Commission, processes thousands of invoices annually from suppliers based in the UK, US, Ireland, and other countries. Each invoice demands structured extraction of supplier name, address, country, and total expenditure, with conventions that vary by supplier country and source document. The corrections the agent must learn are the organization-specific conventions that turn a correct extraction into a usable one.

We evaluated four configurations on a dataset of 100 invoices across 10 independent runs, reporting the average score across runs with 95% confidence intervals. The primary metric is F1 score, balancing precision (are the predictions correct?) with recall (is the model predicting all the expected fields?). Our quality bar is strict: did the user save exactly what the agent predicted, with zero edits? By that measure, a prediction of “UK” when the organization’s records use “United Kingdom” is a quality gap, because the user still had to correct it.

The results

These results are from pre-rollout offline simulation on UK Electoral Commission dataset. The data entry tool extracts content from invoices reliably. The gap is between that raw extraction and the format and conventions the organization expects. Closed-loop learning bridges that gap, calibrating the agent to the business. Across 4277 field instances, closed-loop learning decreased the share of fields users had to edit from 64% to 48%: 1045 fewer fields requiring manual correction.

In 10 independent runs, Genetic-Pareto improved F1 score over the baseline from 66.4% to 74.6%, an 8.2 percentage point improvement.

The optimized prompt outperformed the baseline in all 10 runs, with non-overlapping confidence intervals and a positive confidence interval for the difference.

How Genetic-Pareto calibrated the agent

The F1 number tells part of the story. The baseline typically extracts the right information from the invoice; the gap is in how the business needs that information presented. A supplier name shown as a brand tag rather than the legal entity. An address combined into one line rather than splitting across the fields the organization expects. A total reflecting the ex-VAT amount rather than the gross invoice total.

In one sample run, Genetic-Pareto addressed 76 of 583 gaps in the baseline (a 13% reduction):

These are taste and preference gaps. Genetic-Pareto closes them by calibrating the agent to how the organization structures its data. The gains were most pronounced when specific corrections mapped to a generalized principle. Country accuracy alone jumped from 11% to 78% after Genetic-Pareto learned to expand abbreviated country names. A broader evaluation of additional customers is in progress to validate and expand these findings.

Closing the loop

This evaluation and rollout loop a data science team typically runs happens automatically inside your tenant, through the Power Apps MCP server. The system generates a candidate prompt, runs a shadow experiment (each request uses the current prompt for the user-facing result while the same input is scored in parallel on the candidate), and uses statistical validation (hypothesis testing and power analysis) to decide whether the candidate is measurably better. When it clears the threshold, the candidate automatically becomes the new baseline, and every subsequent request runs on the improved version.

What comes next

Closed-loop learning will extend across more agentic workflows on the Power Apps MCP server over the coming weeks.

The gap between passing the driving test and feeling confident behind the wheel is experience. Closed-loop learning gives agents that experience.

Get started: Add the data entry tool to your agent on the Power Apps MCP server and give your agent a memory.

The post Power Apps MCP server introduces closed-loop learning for enterprise agents appeared first on Microsoft Power Platform Blog.

Here’s what’s new for each persona: 

• For business users, business data in Microsoft 365 Copilot. Your unified business data includes Dynamics 365 or custom apps built in Power Platform powered by Dataverse combined with Microsoft 365 data (i.e. emails, meetings, and documents) to deliver grounded, contextual answers.  

• For makers, business skills describe a specific process. Skills are the detailed steps-by-step instructions involved, the information required, and the business rules that apply. Agents connected to the Dataverse MCP server discover relevant skills automatically and use them to complete tasks according to your organization’s standards.  

• For developers, Dataverse Plugin for coding agents. Dataverse plugin is an open-source plugin that lets AI coding agents like Claude Code and GitHub Copilot build and manage Microsoft Dataverse solutions through natural language. Available on Claude Marketplace: install a plugin, say “connect to Dataverse,” and start building. 

Here’s what we shipped and why it matters. 

Dataverse in Microsoft 365 Copilot

Business data has been locked behind app-specific experiences. Business users context switch between Dynamics 365, Power Apps, Outlook, and Teams to piece together the full picture. What if Copilot could understand your business data the way your best people do? 

That’s now happening. In March, we announced Microsoft 365 Copilot embedded as an in-app sidecar experience within Power Apps, Dynamics 365 Sales, and Dynamics 365 Customer Service. Under the hood, we introduced a reasoning layer natively integrated in Microsoft 365 Copilot, reconciling enterprise data and work signals from across your apps to power the combination of critical sources of business data and insights: Microsoft 365 apps, Dynamics 365 apps, and Power Platform. Coming soon in early June, business data will be available across Microsoft 365 Copilot experiences in Copilot App on desktop, Teams, Outlook, Word, Excel, and PowerPoint.

What does this look like in practice? Ask Microsoft 365 Copilot “There’s new guidance about vendor selection. For overdue issue, please select which vendor to assign” and get a precise list grounded in your CRM records and email signals. No report building. No app switching. No guesswork. Let’s see this in action. 

By leveraging the intelligent semantic layer in Agentic AI powered Dataverse Search, Microsoft 365 Copilot experiences deliver answers using an adaptive reasoning process based on schema and keywords. This allows Microsoft 365 Copilot to understand tables and relationships and how to navigate them to get the right answers rather than doing just keyword search. 

The same proven unified semantic search index that powers global search in Power Apps also provides retrieval and grounding for Copilot, agents, and MCP tools. When your search gets better, every AI experience built on it gets better too. In this release, we’ve made the index faster, more real-time, and easier to manage: 

• Up to 6× faster initialization. Turning on Dataverse search for a new environment now takes minutes, not hours. Your team gets started faster, so they can build more.  
• Near real-time data freshness. Newly added or updated records appear in Copilot results within minutes. Copilot and agents are ground on current business data, not stale snapshots. 
• Zero-disruption schema evolution. Adding or removing columns no longer pauses indexing. Schema changes refresh in the background while Dataverse Search keeps running. 
• Visibility and control for admins. See exactly which tables are indexed, how much capacity each consumes, and download usage reports. Enable or disable indexing for Copilot and search independently, so you can scale AI workloads without impacting existing search behavior. 

One search index for Copilot, agents, and MCP. All grounded in the same governed business data in Dataverse. 

Business skills: teach agents how your organization works 

Your organization’s most valuable knowledge lives as tribal knowledge in people’s heads: the escalation path for a vendor issue, the approval steps for a sales proposal, the process to assign the right specialist to staff your project. Agents can’t access what isn’t simply written down. 

Business skills in Dataverse, now in public preview, captures your processes, policies, and domain expertise as natural-language instructions. Makers can write business skills in plain language or upload existing skills and govern them with built-in controls. The best part: any agent connected to the Dataverse MCP server discovers your skills automatically, whether it runs in Copilot Studio, GitHub Copilot, Azure AI Foundry, or any MCP-compatible client. Define once, apply everywhere – let’s see this in action.  

Velrada makes this real in a field service inspection scenario: 

“Velrada built Inspection Agent to help worksite supervisors and field workers track the maintenance status of their equipment – so they can trust the tools used to get the job done. The Inspection Agent uses business skills to perform an equipment inspection with the user. For an onsite HVAC inspection, the Inspection Agent will invoke the business skill to determine the questions to ask based on equipment class, checks the last inspection outcomes, and pull in context of any historic issues. The result: a conversational assessment that produces a consolidated inspection report on the HVAC unit’s condition and follow-up guidance for maintenance” 

— Matthew Pontel, General Manager of Applied AI, Velrada 

Learn more about business skills: Introducing business skills: Teach agents how your organization works  – Microsoft Power Platform Blog 

Dataverse plugin: coding agents now speak Dataverse 

Enterprise development is shifting from writing code to directing AI agents. Developers describe intent; coding agents orchestrate the right tools over governed business data. What used to require juggling APIs, CLIs, and documentation can now be expressed as a single prompt. 

The Dataverse Plugin for coding agents (in public preview) makes this real. Install one open-source plugin, and your coding agent gets full Dataverse fluency. Under the hood, the plugin packages four tools and knows which one to reach for:  

• Dataverse MCP Servers for ad-hoc discovery and natural-language queries. 

• Dataverse CLI (in preview) for rich data plane actions. 

• Python SDK for batch and scripted operations. 

• PAC CLI for admin gestures like solution export and environment management. 

No manual setup. Describe your intent; the agent handles the orchestration. See this in action. 

Picture this: a platform developer needs to build a customer escalation tracker on Dynamics. With the Dataverse plugin for coding agents, they connect to their environment, create tables, define business skills, configure security roles, and deploy the solution, all through natural language in their coding agent. That’s modern development with coding agents. 

Get started: Install the Dataverse Plugin 

Dataverse agentic evolution: from experimentation to execution 

Six months ago, agents could access your data. Today, they understand your business: your schema, your processes, your rules. The wall between data access and business understanding is coming down: that’s the agent data platform. Today at EU Biz Apps Summit in Cologne, we’re showing all of this live. And we’re just getting started. At //Build in June, we’ll share what’s next. Learn more:  

• Business Skills blog: http://approjects.co.za/?big=en-us/power-platform/blog/2026/05/01/business-skills/ 

• Sample Business Skills on Github: https://github.com/microsoft/dataverse-business-skills

• Dataverse + M365 Copilot blog: Build business understanding with Dataverse in Microsoft 365 Copilot – Microsoft 365 Developer Blog 

• Dataverse docs: https://learn.microsoft.com/en-us/power-apps/maker/data-platform/data-platform-intro 

• Dataverse MCP preview docs: https://learn.microsoft.com/en-us/power-apps/maker/data-platform/data-platform-mcp-preview-tools 

• Dataverse Plugin blog: https://devblogs.microsoft.com/powerplatform/dataverse-plugin-claude-marketplace/ 

• Dataverse Search docs: http://approjects.co.za/?big=en-us/dynamics-365/blog/it-professional/2025/12/08/agentic-ai-dataverse-search/ 

The post Dataverse Is Your Agent Data Platform: Here’s What’s New  appeared first on Microsoft Power Platform Blog.

Then, customers like you pushed us further.

You told us where agents didn’t scale, where authentication slowed runs, and where it was hard to understand why something failed—or to prove it behaved correctly. You also told us where your organization needed more control, visibility, and flexibility before rolling out CUAs broadly.

Today’s updates are a direct response to that feedback.

Computer‑using agents in Microsoft Copilot Studio now offer more model choice, stronger security and governance, and easier scale—so you can automate more of your work across web and desktop apps with confidence.

Here’s what’s new—and why it matters.

Choose the right model to navigate dynamic interfaces

Computer-using agents now support multiple foundation models, including Anthropic’s Claude Sonnet 4.5 alongside OpenAI’s Computer-Using Agent. This gives you the flexibility to choose the best fit for each agent, based on the interface and the task.

• Use OpenAI Computer-Using Agent to orchestrate multi‑step web and desktop flows.
• Opt for Anthropic Claude Sonnet 4.5 when you need high performance reasoning on dynamic user interfaces (UIs) and interpretation of dense, changing dashboards.

Secure authentication with built in credentials and Azure Key Vault

Authentication shouldn’t be the reason automations stall. Computer use now offers built‑in credentials so agents can:

• Securely perform website and desktop app logins
• Reuse them across multiple agents and automations
• Eliminate manual login prompts during runs, enabling unattended execution

For example, if an agent needs to log into a vendor portal and update a desktop ERP every night, built-in credentials now let the agent authenticate to both the web portal and the desktop app automatically. This removes manual interruptions and makes overnight processing dependable while maintaining governance controls. No need to babysit “unattended” runs.

You can choose between two storage options aligned to your governance needs: internal storage (encrypted in Microsoft Power Platform) for low-friction setup, or Azure Key Vault for enterprise-grade secret management.

Credentials are encrypted and are never exposed to the AI model, so only authorized agents can access them. This way, your security and compliance team can feel confident scaling CUAs to more scenarios.

See every computer-using agent action with session replay and audit logs

As agents touch more business‑critical systems, teams need to know what happened, why it happened, and where.

Computer use now has advanced monitoring and richer observability, so operations, security, and compliance teams can inspect behavior step‑by‑step. This includes:

• Session replay with screenshots
• Step‑by‑step action logs (with action types, coordinates, timestamps, and context)
• Run summaries (instruction text, duration, action counts, average time per action, and human escalation counts)
• Resource tracking (including websites, desktop apps, credentials used)
• Export options for offline review

But what does this look like in practice? Imagine an agent run produces an unexpected update, and your team can’t tell whether the agent misread the UI, clicked the wrong control, or encountered a hidden pop‑up.

Session replay and action logs now show exactly what the agent saw and did, pinpoint the step where the UI changed, and produce an exportable record for audit review. That way, you can fix issues faster and retain a defensible compliance trail.

Beyond the monitoring pane, compliance is further strengthened through:

• Microsoft Purview integration, sending audit logs to Purview
• Dataverse logging with configurable verbosity—choose All data, Data without screenshots, or Minimal
• Retention options from 7 days to indefinite, to match regulatory and governance requirements

Simplify infrastructure with managed Cloud PCs for computer-using agents

Scaling UI automation shouldn’t require managing fleets of desktops or fragile virtual machines. The new Cloud PC pool, powered by Windows 365 for Agents, provides fully managed cloud‑hosted machines that are Microsoft Entra joined and Intune enrolled, designed for computer use runs and built to scale with demand.

In other words, these Cloud PC pools provide managed capacity for high-volume runs when demand spikes—without the overhead of keeping dedicated hardware patched, available, and idle the rest of the time. This way, your team can handle spikes without over-provisioning hardware.

Note: For evaluation, you can create up to two Cloud PC pools per tenant with 50 hours of free usage for published autonomous agents—making it easier to pilot CUAs at scale before broader rollout.

Extend—don’t replace—your automation

If you’ve built automations with Microsoft Power Automate and RPA, computer use expands what you can automate—especially when:

• Interfaces change frequently
• APIs aren’t available
• Decision logic becomes more complex

Thankfully, you can keep classic RPA for deterministic scenarios with stable interfaces. CUAs then add flexibility and adaptive reasoning where RPA falls short (such as dynamic web apps, shifting layouts, or complex decisioning). After all, the goal isn’t to start over—it’s to modernize and extend what you already have.

For example, say you have an RPA bot that depends on fixed selectors. Historically, it broke each time a web form changed, forcing constant script updates.

Now, the RPA stays the same, while a CUA handles the variable UI portions—navigating changing layouts, interpreting dialogs, and escalating edge cases. The result? Reduced maintenance and improved reliability.

Get started and help shape what comes next

Ready to try computer‑using agents in a US‑based Copilot Studio environment?

1. Create or open an agent in Microsoft Copilot Studio.
2. Go to Tools → Add tool → New tool and select computer use.
3. Describe the task you want the agent to perform in natural language.
4. (Optional) Choose a model, configure built‑in credentials, and set up a Cloud PC pool for secure, scalable runs.

For deeper guidance, configuration details, and best practices, see the computer use documentation.

Before you go: We’re actively investing in advanced governance, operations, and scale for CUAs—and customer feedback directly informs the roadmap. Tell us what you think of the latest CUA updates today:

• Email feedback to computeruse-feedback@microsoft.com
• Join the Copilot Studio community

The post Computer-using agents now deliver more secure UI automation at scale appeared first on Microsoft Power Platform Blog.

Using Copilot Studio, organizations around the world can quickly bring the benefits of AI to their business. Copilot Studio empowers companies to streamline and automate their processes with agentic workflows, create single-purpose agents to solve specific problems, and develop multi-agent solutions that drive measurable business outcomes at scale. The result: a scalable, secure, and governable foundation that supports the needs of IT administrators and business owners measuring return on investment (ROI). This system accelerates agentic transformation by delivering speed-to-value without sacrificing quality or control.

Microsoft Copilot Studio

Discover practical resources, use cases, and community support to adopt and maximize Copilot Studio.

Get adoption resources

At the same time, with Microsoft 365 Copilot, users can easily use AI to improve their personal and team productivity. This tailored experience for Microsoft 365 Copilot users offers a fast, guided way to set up agents to support your work and automate everyday tasks, removing them from your plate.

Today, we’re excited to share new capabilities in Copilot Studio that support all of these scenarios and groups that use our product, making it easier for makers and administrators to shape agent behavior, enforce organizational standards, and extend functionality with AI.

End-user improvements

Our Copilot Studio experience for building agents and workflows, as well as our agent building capabilities in Microsoft 365 Copilot, continue to support agent creation for all users, from professional makers and IT administrators doing enterprise AI transformation, to employees building agents and workflows for their personal use. Recent updates focus on making the process simpler and more efficient.

What’s new in Microsoft 365 Copilot

• Redesigned creation experience: Build and refine agents through an improved conversational interface that guides users and taps into an expanded set of work-related knowledge sources.
• File generation with natural language: Agents built in Microsoft 365 Copilot, can now create Word, Excel, and PowerPoint files in seconds using natural language commands.
• Seamless upgrade path: Copy agents from Microsoft 365 Copilot to Copilot Studio in one click, unlocking advanced AI agent customization.
• Workflows agent in Microsoft 365 Copilot: Create, build, and manage workflows using natural language in chat. Boost productivity with quick scenarios like daily triage, weekly digests, and lightweight approvals—all directly within Copilot.

Start building workflows with Copilot today

Maker improvements

IT application developers and other professional makers in the business can already build sophisticated agents in Copilot Studio without needing to code. Copilot Studio includes capabilities such as connecting and acting across more than 1,400 systems of record via Model Context Protocol (MCP), Power Platform connectors, and the Microsoft Graph. It also includes broad and deep tooling like autonomously writing and executing code, delivering rich out-of-the-box agent analytics and ROI measurement, and more, all built on the Microsoft governance and security platform. We’re excited to share new capabilities that give makers even more flexibility and control to design enterprise agents tailored to their unique organizational needs.

• Choose your own model: Select from leading options like OpenAI’s GPT‑5, Anthropic’s Sonnet 4.5, and Opus 4.1 to power your agents. This empowers you to tailor agent intelligence to fit your specific business scenario, optimize performance, experiment with new capabilities, and deliver agents that meet your organization’s unique needs.
• Ensure agents are ready for launch, and don’t regress over time, with Evaluations: Built-in evaluation tools help you test agents against real-world scenarios, compare versions, and track performance with clear metrics. Evaluations can give teams greater confidence that their investments are performing as expected.
• Computer use: Agents can now automate tasks across apps and websites, using secure Windows 365 experiences—from hosted browsers for quick web automation to IT-managed Cloud PC pools for rapid scalability.

Admin improvements

As agents become central to automating work and transforming workflows, Copilot Studio is introducing new governance and protection capabilities designed to help organizations maintain strong oversight.

• Expanded agent analytics: Clear insights into connected and child agent performance, detailed visibility into Copilot Credits consumption and limits, AI-generated summaries of top analytics insights, and interrogating analytics using natural language.
• Real-time protection: Copilot Studio integrates with Microsoft Defender and other trusted security platforms, providing continuous monitoring and protection against threats like prompt injection—helping every agent run more safely.
• Microsoft Entra Agent ID: Every agent made in Copilot Studio now gets a unique Microsoft Entra Agent ID, making it simple to register, manage, and govern your entire agent fleet.

Agent 365 and Copilot Studio: Unified control for agents

Agents are handling more responsibilities across enterprise operations and Copilot Studio is your launchpad for building them. With the introduction of Agent 365—the control plane for agents, the rich governance and management capabilities we offer today including sharing controls, advanced connector policies, agent inventory, zoned environment management, and more, will also be surfaced in the Agent 365 platform when using agents built in Copilot Studio.

Learn more about Agent 365

Additionally, in Copilot Studio, makers can now build agents that use the new Agent 365 MCP servers. These servers allow agents to schedule meetings in Microsoft Teams, draft documents in Word, send emails in Outlook, and update customer relationship management (CRM) records in Microsoft Dynamics 365. This supports delivery of intelligent, compliant workflows and agents with built-in audit trails and granular policy enforcement—all from one platform.

Agent 365 is available starting today in Microsoft 365 Admin Center with Frontier, Microsoft’s early access program for the latest AI innovations.

Scale to the Frontier Firm with control

True transformation happens when agents are built for scale, governed for compliance, and measured for impact. Copilot Studio delivers that foundation, so organizations can build enterprise multi-agent systems, automate workflows with precision, and reimagine processes while minimizing risk.

EY’s results show what’s possible when you invest in a comprehensive agent platform, built on Microsoft. They are just one of many enterprise organizations implementing agents with Copilot Studio. In this case, their PowerPost Agent built on Copilot Studio led to major improvements in journal processing:

• 95% reduction in lead time
• 37% cost savings¹

Implementing the PowerPost Agent with Microsoft Copilot Studio has fundamentally transformed our journal processing. What once took minutes now happens in seconds—streamlining workflows and improving accuracy.

—Paula Korczak, Product Manager for General Ledger, EY

That’s the difference between cobbling together siloed agent platforms versus investing in a managed scalable agent platform like Copilot Studio: agents and agented process design that is repeatable, auditable, and scalable.

Get started today

To learn more about Copilot Studio and how it can transform your organization’s productivity, visit the Copilot Studio website and sign up for a free trial today. Take the Agent Readiness Assessment to benchmark your organization’s agent maturity across five critical areas—strategy, data, process, culture, and security—and get a personalized report to accelerate scalable agent adoption and drive agentic business transformation.

Try Copilot Studio for free

Want to explore all of Copilot Studio’s adoption content? Visit the Copilot Studio adoption page.

¹ EY redesigns its global finance process with Microsoft Power Platform

The post Why Microsoft Copilot Studio is the foundation for agentic business transformation appeared first on Microsoft Power Platform Blog.

Today, we are announcing the public preview of Agent Evaluation in Microsoft Copilot Studio, bringing rigor directly into the agent-building tool you already use, backed by Microsoft’s end-to-end approach.

Learn how to evaluate the performance of your agents

Introducing Agent Evaluation

Agent Evaluation enables structured, automated testing directly in Copilot Studio, providing makers with a direct and seamless way to create evaluation sets, choose test methods, define success measures for the agent, and then run the test—maximizing the power of model choice that Copilot Studio offers by evaluating agent performance across multiple agent-level models.  

Create evaluation sets

Makers can now upload predefined test sets, reuse recent Test Pane interactions, and add test questions manually. We are also enabling AI-powered generation of test queries from the agent’s metadata, knowledge sources, and more—delivering makers with quick visibility into agent quality without requiring the manual work for expected answers. This allows for early testing, while additional Q&A sets can be manually added for deeper evaluation. 

Makers can also mix AI-generated queries with manual or imported test sets to expand coverage, helping to evaluate both breadth (common scenarios auto-generated by AI) and depth (organization-specific queries) of agent behavior.

Choose flexible test methods

Makers can choose from a wide rage of test methods—whether it is exact or partial matches, advanced similarity metrics, intent recognition, or relevance and completeness, makers can choose the test methods that work for them based on the type of agent they are deploying. This allows makers to mimic how different users judge the agent—from strict checklist compliance to overall helpfulness—giving a comprehensive view of performance.

Define measures of agent success  

Agent Evaluations allows you to define what constitutes success for your business, whether it is strict keyword matches (lexical alignment) or conceptual, meaning-based matches (semantic alignment). You can also set custom thresholds to ensure your agent meets your organization’s unique standards for accuracy and relevance.

Execute evaluations

Once the dataset is prepared, test methods are chosen, and thresholds are configured, evaluations are executed with a single click. Results are displayed with clear pass or fail indicators, numeric scores on answer quality, and details around the knowledge sources used by the agent. No more guessing as to why an answer failed.

Transforming agent quality: From build to continuous improvement 

Agent Evaluation transforms agent development into a full lifecycle of build, test, and improve. We want makers to have the same rigorous and streamlined quality process for agents as they do for traditional software. By launching evaluations in Copilot Studio, we’re ensuring that every agent can be tested and continuously improved, leading to well-tested agents deployed across the organization. This also enables makers to test agents using different agent-level models for agent orchestration, to find the model that best suits the business process being transformed. You can go from building an agent to testing it in the same interface, all while being confident in Microsoft enterprise-grade permission controls, compliance, and governance capabilities.

Next steps 

To learn how to get started, visit Agent Evaluation in Copilot Studio. 

Check out all the updates live as we ship them, as well as new features released in the next few months here: What’s new in Microsoft Copilot Studio. 

To learn more about Copilot Studio and how it can transform your organization’s productivity, visit the Copilot Studio website or sign up for our free trial today.

We look forward to sharing more about Agent Evaluation at the Power Platform Community Conference 2025.

Build and customize an agent that works for you today

The post Build smarter, test smarter: Agent Evaluation in Microsoft Copilot Studio appeared first on Microsoft Power Platform Blog.

Try Copilot Studio

One unique benefit of using the GPT-5 system is that it enables your agent to utilize your instructions and prompts as a guide for selecting the ideal model using GPT-5’s real-time router for each task or response.

• For simple or routine prompts and instructions, GPT-5’s high-throughput model for chat is used, prioritizing speed and straightforward responses.
• For complex or more open-ended agent steps, GPT-5’s deeper reasoning model will be used, taking its time to craft a plan, gather and comprehend all relevant context, and check its work before providing a thorough response.

GPT-5 to orchestrate your agents and summarize answers 

An agent’s orchestration model powers how it interprets instructions, generates a plan, uses tools, and responds with generative answers. Smarter orchestration means sharper agents. With GPT-5, agents built in Copilot Studio have sharper reasoning in ambiguous situations to determine the right next step, increased coherence in multi-turn conversations, and stronger instruction following, all of which improve overall agent performance. 

From today, makers building agents in early release cycle environments will see a new option to pick their agent’s model in the settings page. If makers select GPT-5 Auto, then the agent will use GPT-5’s real-time router to choose the right model, either GPT-5’s high-throughput model or the deep reasoning model, allowing the agent to self-select the right balance of speed and planning. Alternatively, makers can select GPT-5 Reasoning to ensure that the agent primarily uses the deep reasoning model for complex scenarios. 

GPT-5 to power your agent’s tools  

Tools are a key ingredient in enabling agents to address business processes, and custom prompts are one of the most flexible tools in Copilot Studio’s agent-building toolbox. Custom prompts let makers apply AI models with specific instructions to complete sub-tasks like creating a document in a specific format based on all the information gathered so far or extracting an invoice number from many uploaded invoices. 

Now, makers creating custom prompt tools for agents in early release cycle environments will see GPT-5’s high-throughput model and the deep reasoning models as options to use. With longer context support and improved document understanding, custom prompts using the high-throughput model will offer natural, context-aware responses with faster response times. Using the multi-step logic of GPT-5’s deep reasoning model, prompts complete complex tasks like planning and analysis with increased accuracy. 

Get started today 

With GPT-5 in Copilot Studio, not only are you able to use a cutting-edge AI system, but you also benefit from Copilot Studio’s ongoing optimizations, like agent-specific fine tuning, GPU infrastructure management on a global scale, and simple ways to upgrade existing agents to new models, in addition to having the flexibility of being able to bring your fine-tuned models hosted in the Azure AI Foundry Models.  

GPT-5 in Copilot Studio is releasing today as an “experimental” AI system for use in agents built in early release cycle environments, enabling makers to test in new and existing agents. We anticipate moving GPT-5 in Copilot Studio to all environments later this month. Learn more about early release cycle environments and agent model selection. Read more about GPT-5 in Microsoft 365 Copilot.

Try Copilot Studio

Microsoft Copilot Studio is an intuitive agent development platform, where users can build agents using natural language. It is a secure, enterprise-ready SaaS platform, uniquely integrated with both the Microsoft ecosystem, including Copilot, and external data vendors. Copilot Studio has a complete set of tools that lets you accelerate building the right agents for your business, helping to multiply capacity, reduce development costs, and increase customer satisfaction—fast. 

The post Available today: GPT-5 in Microsoft Copilot Studio appeared first on Microsoft Power Platform Blog.

Microsoft Power Platform enables organizations to build low-code apps and automation within your established governance, controls, and operational models. This framework can be applied to AI agents as well. Microsoft Copilot Studio is built upon the developments and experiences from Power Platform, allowing organizations to utilize their existing resources.

Copilot Studio plays a leading role in the agent shift. According to Microsoft’s FY25 Q3 earnings release, Copilot Studio has been used by over 230,000 organizations, including 90% of the Fortune 500¹. IDC project 1.3 billion AI agents by 2028². The scale and speed of adoption make one thing clear: governance is emerging as a critical priority.

CIOs should consider these five key areas:

1. A Governance Mindset Is Essential for Agents
2. Low-code Lessons Apply Directly to Agents
3. Driving Visibility, Cost Control, and Business Value
4. Empower Innovation with Guardrails
5. Community, Training, and Experimentation Drive Adoption

1. A Governance Mindset Is Essential for Agents

Agents don’t just respond to prompts. They initiate actions and operate across disparate systems. That means governance can’t be static. It must evolve to cover the growing agent behaviors and industry requirements.

Begin by considering agents as digital labor. Assign them trackable identities, define their roles and permissions, and continuously monitor their behaviour and performance.

According to Microsoft’s 2025 Work Trend Index, Frontier Firms—organizations powered by intelligence on tap and human-agent teams—are emerging through three phases of AI evolution: from assistants to digital colleagues running entire businesses processes. This progression is redefining collaboration, as humans shift from users to orchestrators of digital labor.

Not every agent should have the same level of autonomy. Some might only perform low risk activities like answering questions. Others, like a sales development agent, might handle RFPs and pricing proposals autonomously. CIOs should define tiers of autonomy and enforce them with technical guardrails. Just like you wouldn’t give a new hire full system access on day one, agents also need scoped permission and supervision. Consider the agent supervision across review, monitor and protect depending on the agent complexity.

Reviewers are responsible for identifying and reviewing AI-generated output and content to verify accuracy. Monitors observe and track the actions of AI and agents, enabling human or AI-based follow-up as necessary. Protectors have the ability to adjust or restrict AI and agent actions and permissions.

2. Low-code Lessons Apply Directly to Agents

If you have experience with Power Platform, you’re already familiar with this process. You can apply the same playbook: establishing a center of excellence, enforcing security measures like Data Loss Prevention policies, managed environments, and role-based access controls to agents as well.

Maintain consistency by applying your existing compliance, security, and audit frameworks to agents, updating them for new behaviors as needed. In addition to using Power Platform Admin Center, leverage other Microsoft tools like Purview and Entra ID, and ensure your governance framework supports safe innovation.

Additional IT guidance content can be found on the adoption site.

3. Driving Visibility, Cost Control, and Business Value

Visibility is the foundation of effective agent governance. Without it, agents can proliferate unchecked, leading to redundancy, security gaps, and unnecessary costs. This is why CIOs must establish reliable telemetry that offers deep insight into who created an agent, what data it accesses, how often it’s used, and the resulting impact on the organization’s resources.

Fortunately, tools like Copilot Studio’s built-in analytics and Power Platform Admin Center offer the transparency and insights to manage agent usage and costs effectively. By tracking consumption and reviewing performance regularly, teams can identify underused or redundant agents, forecast expenses with tools like the cost calculator, and ensure agents stay aligned with strategic goals.

Read the agents cost management E-book here

While managing costs helps keep investments in check, it’s the business value that ultimately justifies them. CIOs should look beyond usage limits and budget forecasts to ask a more strategic question: what outcomes are agents actually driving? This shifts the focus from spend to impact.

Ultimately, governance without visibility is just guesswork. Robust telemetry ensures that every agent is accounted for, managed wisely, and contributing to safe, scalable innovation.

4. Empower Innovation with Guardrails

The people closest to the work often have the best ideas for how agents can help them. Empowering business teams to build their own agents can accelerate innovation and speed.

But empowerment without guardrails is a risk. All agents must operate within strict security and compliance boundaries. Enforce permission models so agents only access authorized data sources. Use environment strategies and connector policies to keep sensitive data safe and audit each key step.

A zoned governance model, with centralized policy and progressive autonomy, gives CIOs a scalable way to manage agents. IT sets boundaries allowing business units to innovate safely within these zones:

• Zone One: Personal Productivity – The entry point for experimentation and innovation provides isolated environments where individuals can safely explore agent capabilities, guided by governance and security policies.
• Zone Two: Collaboration – This zone supports team-based agent development with stronger controls, including environment-level policies, connector restrictions, and operational oversight. It enables broader adoption while maintaining compliance and consistency.
• Zone Three: Enterprise Managed – The most advanced zone, designed for production-grade agents. It includes enhanced security protocols, continuous monitoring, and structured lifecycle management. This zone supports complex, cross-functional and autonomous agent scenarios with full visibility, scalability, and strategic alignment.

Scaling agent deployment effectively requires not just the right tools, but also thoughtful organizational structures and clear assignment of roles and responsibilities. Establishing rhythms and governance frameworks ensures responsible agent management across the organization.

As organizations operationalize agents and build the structures to support them, CIOs will likely encounter demand for roles that didn’t exist just a few years ago. They’re emerging in response to the unique demands of building, governing, scaling, and securing AI and agent systems responsibly.

5. Community, Training, and Experimentation Drive Adoption

People are the engine behind every successful technology initiative—and AI and agents are no exception. The biggest challenges in agent governance aren’t technical. They’re cultural. To succeed, you need more than policies and platforms. You need people who are bought in, equipped, and empowered. 

Build an active agent community hosting events such as “Agent Show-and-Tell” and hackathons. Acknowledge successful projects and appoint departmental champions to mentor others and drive adoption.

Training should cover both agent development and guidance on responsible governance. Support users with learning paths based on their different AI readiness levels and take advantage of the agent creator community.

Support experimentation within a structured framework. The Center of Excellence should manage best practices, training, and governance, gathering insights to improve and scale effective approaches.

What to Do Next

CIOs are uniquely positioned to lead the agent transformation by building and evolving on what already works. The governance models, CoEs, and controls you’ve established for Power Platform don’t need to be reinvented, they need to be extended to incorporate agent autonomy, decision making and responsible AI.

Calls to action:

1. Governance is the foundation, not the finish line.

Agents introduce new opportunities but also risks and responsibilities. CIOs must lead with a governance mindset that treats agents like digital labor—assigning identities, defining autonomy, and enforcing oversight through familiar tools like PPAC, DLP, Purview, and Entra ID.

2. Culture will make or break your agent strategy.
Technology alone won’t drive adoption. Build a community of practice, empower champions, and invest in training that reinforces not just how to build agents—but how to govern them responsibly.

3. Ready to operationalize? Start here.
Download the e-book for detailed insights and a shareable copy of the five sections.

Check out these additional resources to get started

• Agent Governance Whitepaper
• Implementation Guide
• Agent Success Kit
• Agents Cost Management E-Book
• Agent Creator Community
• How to deploy transformational enterprise-wide agents: Microsoft as Customer Zero

Citations:

1: Microsoft Earnings Release, Call Transcript, FY25, Q3

2: IDC Info Snapshot, sponsored by Microsoft, 1.3 Billion AI Agents by 2028, #US53361825 and May 2025

Disclaimers

This blog is for informational purposes only and does not constitute legal, regulatory, or compliance advice.

The strategies, tools, and governance models referenced herein are based on Microsoft technologies and may not be suitable for all organizations, industries, or jurisdictions.

Any forward-looking statements are subject to change and should not be interpreted as commitments or guarantees.

The post Evolving Power Platform Governance for AI Agents appeared first on Microsoft Power Platform Blog.

We are thrilled to announce a series of recent enhancements to DLP for Power Automate and Copilot Studio, which significantly bolster governance, scalability, and usability across both cloud and desktop automation scenarios.

How does DLP for flows work?

DLP for Power Automate and Copilot Studio is currently generally available for all users. The feature can be accessed from the Power Platform Admin Center, where administrators can create and manage Data policies.

DLP evaluation occurs when editing and saving a flow or when policies are updated.

When a policy change is implemented, flows that do not comply with the Data policy are automatically suspended, and connections to blocked connectors are disabled.

Recent changes that you can now use in your organization

Trigger configuration

Administrators can now block triggers in the same way they can block actions in the DLP configuration experience (triggers are tagged with [TRIGGER]). This provides greater control over which triggers can be used within your organization.

Endpoint Filtering Public Preview for Browser Automation

A new Endpoint Filtering capability is in Public Preview. Endpoint filtering allows administrators to define the websites desktop flows can access by configuring the Browser Automation connector in Data Policies —an essential control for high-risk automation scenarios. Application-level filtering via UI Automation will follow soon this summer.

Environment variable

Use environment variables to migrate application configuration data in solutions. Environment variables are now supported in endpoint definitions, using the @environmentVariables("environmentVariableName") pattern.

For example, an SMTP endpoint can be allowed using @environmentVariables("smtpEndpoint"),587 where the environment variable is defined as smtp-mail.outlook.com.

Recent improvements that make your organization more secured

Runtime Enforcement

We have completed the rollout of support for runtime enforcement as of June 2025. This ensures that Data policies are consistently enforced not only at design time but also during execution.

Automatic reactivation of cloud flows

Policy updates can be disruptive, and administrators may occasionally revert updates that have unexpected effects. We have modified how policy rollout works to now activate compliant cloud flows that had been suspended in the previous 7 days due to policy violations, instead of disabling them. This makes it easier and faster to restore the state of a previous policy version.

Improve performance for large tenants and policies

We have recently strengthened our backend systems to better handle large-scale policy changes with smarter job throttling and improved detection strategies, reducing the time it takes for policy updates to take effect.

The post Announcing major DLP enhancements for Power Automate and Copilot Studio appeared first on Microsoft Power Platform Blog.

The Upwork and Microsoft Power Platform collaboration opens a world of possibilities for organizations eager to leverage innovative low-code solutions, AI Agents, and automation tools. Businesses can now access certified professionals through Upwork’s Power Platform Experts page, helping them accelerate project timelines and unlock new efficiencies. This approach offers flexibility and access to specialized talent, helping organizations stay competitive in a fast-moving digital landscape.

According to Upwork’s Future Workforce Report, freelancers collectively generated $1.5 trillion in earnings in 2024 as independent professionals. For developers, consultants, and solution architects specializing in Microsoft Power Platform, this partnership is an invitation to elevate careers and expand their reach. By signing up to become an Upwork Power Platform Partner Expert, professionals can connect with a global clientele and apply their certifications to real-world projects. While individual results may vary, this program enables participants to contribute to transformative solutions for businesses of all sizes.

Microsoft certifications can serve as a launchpad to freelance opportunities in this program, enabling professionals to contribute to innovative projects while building a flexible and rewarding career.

Redefine what’s possible, join the Power Platform Expert Freelancers to explore new opportunities, supercharge your projects, and innovate globally.

The post Empowering a New Era of Agent-Based Experiences: Upwork and Microsoft Power Platform Partnership appeared first on Microsoft Power Platform Blog.

Microsoft Copilot Studio

Explore new opportunities for applying AI to your unique business needs.

Try for free today

An agent is an AI assistant designed to automate and execute business processes, working with you or for you. They can build capacity for every individual, team, and organization—from sales to marketing to customer service, and more—enabling you to scale impact like never before.

Agents come in all shapes and sizes. They help you retrieve information from grounding data and reason over it to summarize or answer questions. More capable agents take actions when asked, replacing repetitive tasks, and the most advanced agents will be autonomous, operating independently to create plans, orchestrate other agents, and learn to know when to escalate to an employee.

With Copilot Studio, you can create and publish agents to all kinds of channels, including Microsoft Teams, your website, or your customer engagement hubs, and we’ve seen great customer momentum with more than 50,000 organizations having used Copilot Studio to date. 

We’re announcing two new capabilities for Copilot Studio:

• First, agents built in Copilot Studio can now be published directly into Microsoft 365 Copilot, surfacing right in the user’s flow of work.
• Second, we’re unveiling a new Copilot Studio experience that empowers users of any skill level to create agents right from within Microsoft 365 Copilot Business Chat and SharePoint.

These new features will be rolling out in general availability to all Microsoft 365 Copilot customers over the coming weeks. 

Publish Copilot agents to Microsoft 365 Copilot

Copilot agents enable you to enhance Copilot’s capabilities by connecting to your organization’s knowledge and data sources. When Copilot can draw from valuable data in SharePoint, Microsoft Dynamics 365, or your line of business systems, it becomes significantly smarter about your business and its processes. You can also expand Copilot’s capabilities by equipping it with new skills, such as sending emails, updating records, or creating support tickets. And Copilot agents are capable of integrating with your existing line of business systems, enabling them to take action wherever it’s needed.

Users can find agents in a new Copilot agent store in Microsoft 365 Copilot Business Chat, Teams, and the Microsoft 365 app, and they can use these agents by mentioning them (@mention) or by interacting directly with them. IT admins can curate the catalog of agents available in the store from the Microsoft 365 admin center.

Empower users of any skill level to create agents

When it comes to agent creation, we’re excited to introduce a new, lightweight Copilot Studio experience that makes it simple for users of any skill level to build an agent directly within Business Chat and SharePoint. This new experience will roll out in general availability to all customers through October.

With it, you can tell Copilot Studio what you want your agent to do using natural language, and easily choose which knowledge sources you’d like to integrate from across your organization’s data, whether it’s specific documents or entire databases. When you’re ready, the agent can be used by you, shared with a select group, or deployed to an entire organization. Building an agent in SharePoint is even easier and you can create agents in SharePoint with a few clicks, no coding skills required.

You can continue to customize these or any agents in the full Copilot Studio editor, adding more advanced capabilities including autonomy. As we announced at Build 2024, we believe autonomous agents will fundamentally change business processes. We’re actively working with enterprise customers who are building autonomous agents in Copilot Studio as part of our private preview, and we look forward to sharing more later this calendar year about how users can build and use autonomous agents.

Manage agents with centralized controls

We believe organizations that embrace AI will create and use many agents. Copilot Studio provides a trusted environment that delivers confidence through comprehensive policies, access controls, and agent management, all easily managed within the admin center. This includes the ability to control agent creation, sharing, and use, which allows organizations to meet their security, compliance, and governance goals.

Furthermore, the data used to ground agents remains securely within the Microsoft 365 trust boundary, adhering to your organization’s labeling and policy guidelines. When extending your agents through Copilot Studio, your data remains in the Microsoft Azure tenant boundary with security layers to prevent invalid authentication and authorization and data loss prevention (DLP) policies for additional control. And all of this is built on Microsoft’s Responsible AI and Enterprise Data Protection standards.

Ultimately, in a world where AI is rapidly transforming how we work, Microsoft 365 Copilot and Copilot Studio offer new opportunities for applying AI to your unique business needs. By empowering users to create, tailor, and deploy agents of all types, Microsoft is making advanced AI accessible and valuable to everyone.

• Learn more about Microsoft Copilot Studio.
• Learn more about Microsoft Copilot Studio governance. 
• Read the Microsoft 365 blog post about Copilot Wave 2.
• Join us at the Power Platform Community Conference to see how Microsoft is transforming the developer experience. 
• Get in touch with a Microsoft representative today.

The post Unveiling agents built with Microsoft Copilot Studio to supercharge your business appeared first on Microsoft Power Platform Blog.