Build Automated Agents, Workflows, and Apps – New in Copilot Studio
To create your ideal agent, start by describing its purpose and functionality. From there, you can generate specific instructions and reusable skills, select the model, integrate MCP tools along with specialist agents, and finally evaluate its quality through systematic testing.
Set off workflows triggered by incoming emails, categorise submissions, delegate tasks to agents, and pause for human input on critical decisions. Use the Activity tab to track the timing and progression of each step. Plus, you can establish a full-stack application using your Dataverse data, enabling your team to filter, review, and approve all bids from one central location.
Jack Rowbotham, Senior Product Manager at Copilot Studio, highlights how to automate the procurement process for bids from various vendors with intelligent agents that can reason, route, and make recommendations.
Copilot Studio simplifies the development process. It transforms your requirements into an agile agent, carefully outlining the instructions and producing reusable skills as you progress. Learn more here.
Utilize Evaluations in Copilot Studio to conduct thorough tests. You can import your own test cases or let AI create them for you, and compare outcomes as your agent improves. Find out how here.
Apps provide essential structure. Use Copilot Studio to merge all elements and create comprehensive business solutions. Explore this feature.
This tool is fantastic for IT administrators, Power Platform builders, solution architects, developers creating agent-based solutions, and for procurement teams streamlining reviews and approvals.
00:00 Develop agents, workflows, and applications in a single platform
01:04 Streamline a multi-vendor bid review from start to finish
02:18 Create an agent from a prompt
03:28 Incorporate models, MCP tools, and specialised agents
04:13 Test the bid agent in Preview mode
05:05 Evaluations process
05:59 Automate workflows effortlessly
07:42 Monitor every process in the Activity tab
08:13 Build an app from your prompt
09:10 Approve bids directly from your team’s application
09:37 Begin your journey with Copilot Studio
Copilot Studio combines agents, workflows, and applications in one platform, all powered by the GitHub Copilot framework. In this demonstration, an email from a supplier initiates a workflow where a bid evaluation agent checks the submission against the tender specifications, leaving the final decision in the hands of the procurement team via a shared application.
Begin by detailing the agent you wish to create. Specify the assessment criteria, the logic required, and the live data connections needed. Copilot Studio intelligently interprets your prompt, outlines the steps for constructing the agent, and generates its name, along with detailed instructions and skills for requirement coverage and evidence evaluation. Integrate Work IQ for handling supplier emails and connect the Dataverse MCP server for bid data, using Claude Sonnet as your reasoning model, grounding the agent in an RFP example PDF, while adding a Supplier Risk specialist agent to enhance its capabilities.
Test the agent within the Preview tab and observe its reasoning process as it identifies gaps in certification, cold-chain adherence, and insurance requirements. You’ll also want to use Evaluations to check the agent’s performance, whether through your own dialogues, auto-generated tests, or by importing a CSV file of cases. For instance, one evaluation measure showed that 18 test cases met the necessary criteria 89% of the time.
In the Workflows tab, easily integrate steps like an Outlook “When a new email arrives” trigger, a Classify step for RFP assessments, a Get Attachment connector, and an Agent step executing the Bid Evaluation Agent. Introduce a Human Review step for flagged bids needing human oversight, then deploy a Researcher agent and a recommendation agent to conclude the process before results are sent to Dataverse.
Next, build a Bid Management App that links to your Dataverse table or data sources such as SharePoint using available connectors. After branding the app to fit your company’s style, filter for bids requiring human review and make decisions straight from the interface.
You can experience it for yourself at copilotstudio.microsoft.com. Keep updated with Microsoft Mechanics for further enriching content like this.
— Launch your Copilot Studio journey: https://copilotstudio.microsoft.com
Microsoft’s Official Video Series for IT
— Subscribe for more: https://www.youtube.com/c/MicrosoftMechanicsSeries
— Join the Microsoft Tech Community: https://techcommunity.microsoft.com/t5/microsoft-mechanics-blog/bg-p/MicrosoftMechanicsBlog
— Listen to the Podcast: https://microsoftmechanics.libsyn.com/podcast
— Follow us on Twitter: https://twitter.com/MSFTMechanics
— Connect with us on LinkedIn: https://www.linkedin.com/company/microsoft-mechanics/
— Check us out on Instagram: https://www.instagram.com/msftmechanics/
— Explore our TikTok: https://www.tiktok.com/@msftmechanics
– If you’re looking to develop agents that operate beyond a standard chat interface, autonomously carrying out defined workflows while interacting with other agents and tools, and offering new interactive app experiences, I’ll show you how Copilot Studio can help you achieve that and more. You’ll find it’s straightforward. Copilot Studio allows you to craft complete solutions with agents, workflows, and applications all in one place, without requiring a single line of code. You can create individual agents with the model you select, incorporating expert skills, tools, prebuilt connectors, APIs, and integration with your existing systems, all while leveraging specialised knowledge and memory—thanks to the robust GitHub Copilot framework behind it.
– You can connect them with your defined multi-step workflows, ensuring consistency and, if necessary, incorporating skilled agents or human expertise along the way. Finally, you can build comprehensive app experiences using natural language; simply express your requirements and let AI build the app for you. To give you a practical example, I’m going to guide you through creating a solution that manages incoming work bids from various vendors, leading to eventual human approval.
– Here’s the step-by-step breakdown of how it operates in Copilot Studio. First, as a supplier email gets received, it automatically triggers a workflow. This workflow sorts the submission, accesses supporting documents, and submits everything to the bid evaluation agent. This agent reviews the bid against the tender requirements, checks validation of evidence, and flags any potential concerns. If a decision needs human insight, the workflow pauses to involve the appropriate reviewer. A Microsoft Copilot Researcher agent can then gather additional details, while the recommendation agent synthesises the information into a final suggestion. The status information is then passed to a connected data source, allowing our app frontend to provide the procurement team with a unified space to oversee every bid and address any flagged issues before deciding the next steps. Thus, the workflow illustrates each step, decision point, agent handoffs, and results.
– Now, let’s revisit the process of how to create this agentic solution. With several cooperating agents, I’ll kick things off by designing our bid evaluation agent, which will serve as our main reasoning unit for this operation. I’m in Copilot Studio, where I need to simply describe the agent I wish to develop.
– For instance, I’ll instruct it to create a Bid Evaluation Agent tailored for the procurement team. Next, I’ll define how it should assess the bidding criteria, the logic it needs to apply, as well as the live data source to connect to. It must evaluate the proposal requests sent via email and compare the suppliers’ responses to the criteria, delivering a pass or fail verdict along with a final recommendation.
– After submitting this detailed prompt, I let the system process the information. You’ll see that it calculates and will begin constructing the agent following the outlined steps. Additionally, it will automatically create reusable skills to ensure consistency in covering requirements and verifying evidence, which can also be reused across other agents or teams in the future.
– Now that the agent is built, let’s review what it has created. You’ll see a generated agent name and clear instructions. Furthermore, it has established two skills, and you can also upload your skills or develop more using AI just by describing your needs. It has linked to existing tools, specifically Work IQ for Supplier Submission Emails and the Dataverse MCP to manage Procurement and Bid Data.
– You can also add further tools, like MCPs for both Microsoft and non-Microsoft data and services. Plus, you can choose the model that suits your task best. I’ll stick with Claude Sonnet as the reasoning model for our case, grounding it in knowledge sourced from various online services. In this instance, I want to provide it with an RFP example, which I can easily attach from my local device as a PDF.
– Moreover, I can integrate specialist agents, such as one for evaluating Supplier Risk, which I will connect to now. Additionally, there’s the option to retain memory throughout interactions and use that for later sessions. With the main agent now built, I’m prepared to put it to the test in Copilot Studio. I’ll start by using the agent’s Preview tab, pasting an example email I’ve copied and attaching the work bid.
– After that, I just need to submit it to commence the evaluation process. The agent will assess each requirement, verify the related evidence, and pinpoint gaps in compliance. You can actually observe the flow of its reasoning. In this case, it flags that the supplier lacks adequate certification for medicine manufacturing, doesn’t meet the cold-chain handling standards, and falls short on the minimum insurance criteria. This agentic process may take several minutes to complete, so to save time, I’ll jump to the Final Recommendation. Here, it recommends a human review and not to shortlist the supplier until these issues are addressed, providing detailed justifications in the summary.
– That outcome resulted from a single preview run. We can also assess the quality of the agent on a larger scale by using Evaluations in Copilot Studio. Rather than checking sporadically, you can conduct structured tests to measure performance over time. You can draw in your dialogues as test cases or use Copilot Studio to auto-generate tests based on your agent’s design. In my example, I have test cases ready, which I’ll import as a CSV file. This particular file contains 18 test cases with sample requests.
– Once imported, I’ll name the evaluation and initiate it. The system runs these tests sequentially, allowing you to delve into each one for details post-run. The entire batch of 18 tests took just over 22 minutes to complete, and I’ll now check the results, revealing that our evaluations met the criteria 89% of the time. Scrolling through the individual test results displays outcomes for each.
– With our agent functioning, let’s set up the automated workflow that encompasses actions preceding and following the agent’s work. Back in the Workflows tab of Copilot Studio, adding elements is as simple as dragging and dropping onto the canvas. By default, the trigger is manual, but I’ll alter that so it triggers with an incoming email instead. I’ll adjust the trigger type to Connector, selecting Outlook and the “When a new email arrives” option. This can be linked to an individual email account or even to a dedicated RFP intake account, depending on your needs.
– Next, I’ll incorporate a Classify step to assess the email’s subject and message content. Here, I can enter a category, like ‘RFP’, while also having the potential to define multiple categories. Following this, I need to add a step that retrieves the attachment, so I’ll use the Connector option again and search for ‘Get Attachment’. There it is. The attachment is essential for the evaluation process that follows. This leads directly to the Bidding Evaluation Agent that conducts the analysis on the email’s bid attachment.
– I’ll add an Agent step and choose the Bid Evaluation Agent we developed earlier. From this point, I’ll continue adding steps until I’ve completed the workflow. For efficiency, I’ve pre-added the remaining steps. This includes a Human Review step with an if/else decision, allowing the agent to request human intervention when it flags a bid that requires further scrutiny before proceeding. If there are no issues, it will move directly to the Researcher agent to compile more insights about the bid and the bidder.
– The last step in the workflow uses the information gathered from the previous stages to assess the bid and issue a final recommendation. Subsequently, it records a new entry in Dataverse, serving as the backend data source for this process. Each workflow step can be tested independently, making troubleshooting and iterative adjustments straightforward.
– I’ve been sending emails with attached bids to test the process, and can monitor the outcomes of each run. The Activity tab in my workflow displays timing per step and the routes taken for incoming RFP emails. Clicking into any step provides its details; for instance, the agent’s structured output will show whether human review was necessary. After the reviewer takes action, the process continues until the final recommendation gets logged in Dataverse. All these actions are sequenced as an automated background process, while an app will enable team members to view and address details for each bid. Now let’s build that app. This time, I’ll begin in Copilot Studio’s Apps section, where using natural language, I can define the experience I want.
– In this case, I need a Bid Management App to oversee the processing of multiple bids, allowing the team to monitor each bid along with its details. I’ll instruct it to link to my Dataverse table, with the option to connect to other data repositories, such as SharePoint, using available connectors. Below that, I’ll detail the app screens I want to create. I’ll head off with this, which is another extensive process taking several minutes, so I’ll skip to the first version of my app. It’s already looking impressive with summary tiles on top and detailed views of each bid. Using the prompts, I can further refine the app—such as updating its design to match our corporate branding. Within moments, it adjusts the visuals to reflect our company’s style, and it’s ready for testing.
– Indeed, I can switch to the version that other procurement team members would use. I see a full overview of the Bid Queue, including supplier names, statuses, and recommendations. It’s easy to filter bids needing human review, and clicking into any one of these bids displays extra details along with available actions. From here, it’s straightforward to approve a bid, which I’ll do in this instance, and then save it to confirm. Everything functions smoothly, completing the end-to-end process.
– And that’s how agents, workflows, and apps collaborate in Copilot Studio. To get started, visit copilotstudio.microsoft.com and keep watching Microsoft Mechanics for the latest updates in technology. Thank you for your attention.
Share this content:
Discover more from Qureshi
Subscribe to get the latest posts sent to your email.
