Earlier this week, I hosted a Voitanos webinar where I walked through the different options you have for creating custom agents for Microsoft 365 (M365) Copilot. Nearly 200 people joined me live, and based on the questions that came in before, during, and after the session, it’s clear that this is a topic a lot of you are wrestling with!
If you missed it or want to revisit any part of it, the full recording is available: watch the Evaluate Your Microsoft 365 Copilot Agent Options webinar on demand.
Each of the following sections have jump point chapters in the webinar recording in case you want to skip ahead to that topic.
Let me walk you through what we covered and share some of the best Q&A from the session.
Want to Learn More About Building Declarative Agents?
Building off this webinar, I have another one coming up on April 15, where I’ll go deeper into building declarative agents for Microsoft 365 Copilot using the Agents Toolkit in VS Code.
I’ll walk through demos, explain the architecture in more detail, and answer your questions live. Register for the free webinar here.
When you register, there’s a question asking what you want to learn. Answer it, because I’ll use your responses to shape the content, just like I did for this webinar.
Discover the pro-developer path to extending Microsoft 365 Copilot with declarative agents — built in VS Code with the Microsoft 365 Agents Toolkit.
https://www.voitanos.io/webinars/microsoft-365-copilot-developers-guide-declarative-agents-20260415/
Why build a custom Microsoft 365 Copilot agent?
This is the question Microsoft never really answers. They just tell you to start building agents. That doesn’t help.
I started the webinar (12:17) by framing it like this: if you’ve ever used the same prompts over and over, coached Copilot with the same instructions repeatedly, wished it could connect to an external system, or wanted it to take actions on your behalf… that’s why you need an agent. An agent lets you create a repeatable, scenario-specific conversational experience that’s tailored to your business.
Four signals you need a custom Microsoft 365 Copilot agent
I also referenced a Wall Street Journal article from February 2026 (14:13) that reported on a study of over 150,000 Copilot users. The trends weren’t flattering for Microsoft: primary usage dropped from 19% to 11% over six months, many users cited better quality elsewhere, and some companies were only using 10% of their Copilot licenses.
My take?
If you create custom agents that are tailored to your organization, you make Copilot more useful and more “you.” That’s how you drive adoption.
Andrew Connell
Microsoft MVP, Full-Stack Developer & Chief Course Artisan - Voitanos LLC.
Microsoft's Pivotal AI Product Is Running Into Big Problems
Wall Street Journal - February 3, 2026 Sebastian Herrera
How Microsoft 365 Copilot works: prompts, RAG, and the semantic index
Before diving into the agent options, I spent time on the foundational concepts (19:17) because I find that if you don’t understand how Copilot works under the hood, you won’t understand what opportunities you have to customize it.
How Microsoft 365 Copilot combines prompts, RAG, and knowledge sources
We walked through how prompts, system prompts, and RAG (Retrieval-Augmented Generation) work together, and the difference between foundational knowledge (what the model learned during training) and custom knowledge (your organization’s documents, data, and systems). This is what makes agents valuable: they bridge that gap between generic AI and your specific business context.
How a Microsoft 365 Copilot query flows through the orchestrator
The semantic index: Microsoft’s Copilot moat
I walked through the Copilot tech stack (26:27): the user experience, the orchestrator, knowledge sources, skills, and the foundational model. The orchestrator is the component that receives your query, determines what data to retrieve, and coordinates the round trips between knowledge sources and the language model. The key takeaway is that when you ask Copilot a question, it doesn’t just make one trip to the language model. There are at least two round trips, sometimes more, as the orchestrator figures out what data it needs, retrieves it, and then sends a complete prompt to the model.
Then I got into what I consider Microsoft’s moat: the semantic index (34:39). This is the thing that none of the other AI chat bots can do. Microsoft has created a semantic index on top of Microsoft Graph that understands the meaning of your content, not just keywords. It’s what makes Copilot’s responses grounded in your organizational data.
Architecture of the Microsoft 365 semantic index built on Microsoft Graph
Here’s the thing you need to know: the semantic index requires the $30/user/month Microsoft 365 Copilot license ($21/month for organizations under 2,000 employees). If you’re on pay-as-you-go (Microsoft 365 Copilot metered usage) only, you don’t get the semantic index, and that’s a massive gap in what Copilot can do for you.
The four agent creation options
This is where I called out Microsoft’s marketing (40:08). They want you to think you have two options: Copilot Studio or go full pro-code with Azure AI Foundry and build everything yourself. That’s not true. It’s not a binary choice.
You actually have four options for extending Microsoft 365 Copilot, and they sit on a spectrum from simple to powerful. The following comparison reflects capabilities as of April 2026, during the 2026 Wave 1 rollout.
| SharePoint Agents | Agent Builder | Copilot Studio | Declarative Agents (ATK) | |
|---|---|---|---|---|
| Target builder | End users | End users | Power users, IT pros | Developers |
| Knowledge sources | Documents in library | SharePoint, Teams, email, calendar, web | Multiple incl. Dataverse, REST APIs | SharePoint, Graph connectors, REST APIs |
| Actions | None | Limited | REST APIs, Power Automate, triggers | REST APIs, OpenAPI actions |
| Orchestrator | Sydney | Sydney | Samba | Sydney |
| ALM / source control | None | None | Limited | Full (JSON/YAML files, Git, CI/CD) |
| Deployment scope | Library-scoped | Personal / small group | Enterprise | Enterprise (Teams app catalog) |
Four Microsoft 365 Copilot agent options on a spectrum from simple to powerful
From SharePoint agents to declarative agents with the Agents Toolkit
SharePoint agents
Starting at 44:10 in the recording, this is the simplest option. You create an agent inside a document library, give it a name, description, and some instructions, and it has access to the documents in that library. That’s it. No actions, no external connections. Super simple, but very limited.
Agent Builder
Starting at 50:17, this was formerly called Copilot Studio Light. This is the experience you get when you click “create an agent” inside Copilot chat. It gives you more knowledge sources than SharePoint agents: you can pull from SharePoint, Microsoft Teams, email, calendar, contacts, and even web search. You can also add capabilities like image generation. But these agents are really intended for you or a small group, not enterprise-wide deployment.
Copilot Studio
Starting at 53:40, this is the full-featured, low-code, web-based agent builder. This is what Microsoft pitches as their primary agent creation platform. You get a lot more here: multiple knowledge sources, actions via REST APIs, Power Automate flows for workflows, topics for guided conversations, triggers, and the ability to change the underlying model. It’s targeted at power users and information workers who want a visual experience.
Declarative agents with the Microsoft 365 Agents Toolkit
Starting at 59:15, this is the developer-oriented option, but I pushed back on Microsoft’s characterization of it as “pro-code only.” It’s really JSON and text files. You describe what you want the agent to do declaratively: its instructions, knowledge sources, capabilities, and actions. You build it in Visual Studio Code (VS Code) using the Microsoft 365 Agents Toolkit (ATK) extension.
The Venn diagram between what you can do in Copilot Studio versus declarative agents is almost a perfect circle. There are a few things each can do that the other can’t, but there’s massive overlap with a lot of footnotes about how things work differently under the hood.
I showed a demo of my PTO request agent built as a declarative agent, walking through the manifest, the agent definition, capabilities (OneDrive/SharePoint, email, Teams messages, web search, Dataverse tables, Graph connectors), actions, and instruction-based prompt engineering.
How Microsoft 365 Copilot orchestrators differ between agent types
The tool you use to build your Microsoft 365 Copilot agent determines which orchestrator processes all queries, and you cannot change it later. Declarative agents use an orchestrator codenamed “Sydney.” Copilot Studio uses one called “Samba.” The same query against the same content can return ten documents via Sydney and only three via Samba.
This is the part of the webinar (1:10:10) that always gets the biggest reaction, no matter how many times I present it.
Declarative agents and Copilot Studio use different orchestrators
For example, if you build the exact same agent with the exact same instructions, name, and description, the declarative agent version can return ten documents from SharePoint while the Copilot Studio version returns three. Same query, same content, different results. There’s no way to change that today.
Knowledge retrieval differences between Sydney and Samba orchestrators
External connections behave differently too. In Copilot Studio, external connection data gets indexed into Dataverse and is only available within that Copilot Studio experience. With declarative agents, you create connections in the Microsoft 365 admin center, where you have more control over what gets indexed, and that content is available across Microsoft Search and other M365 experiences, not just your agent.
Capability differences between Sydney and Samba orchestrators
Here’s the kicker: Microsoft’s own first-party agents like Researcher and Analyst are built as declarative agents using the Sydney orchestrator. New capabilities land on that orchestrator first, and sometimes Microsoft won’t even announce them at conferences because they’re not in Copilot Studio yet.
How to choose the right Microsoft 365 Copilot agent type
I wrapped up the main content (1:16:12) by walking through the key questions you need to ask when deciding which approach to use. There’s no simple decision tree here; it’s a conversation:
Decision factors for choosing the right Microsoft 365 Copilot agent type
Who is building the agent?
If your team is developer-oriented and values source control and SDLC, lean toward declarative agents with ATK. If you’re more on the low-code/no-code side, Copilot Studio is a better fit.
Who are the end users?
Agent Builder works for personal or small-group use. Copilot Studio and declarative agents support broader enterprise deployment.
What knowledge sources does the agent need?
Some sources aren’t available in every option. Understand what’s supported before you start building.
What actions does the agent need?
Autonomy and triggering capabilities exist in Copilot Studio but not yet in declarative agents.
What are the governance requirements?
If you have strict governance needs, think carefully. I showed a real example from Defcon where someone demonstrated finding publicly exposed Copilot Studio agents that leaked entire Salesforce CRM data without authentication.
How will it be published and shared?
Declarative agents are deployed as Teams apps. Copilot Studio has a more open publishing model, which can be both a feature and a risk.
Microsoft 365 Copilot agent Q&A: licensing, security, and architecture
We had a really active Q&A session, and I want to share some of the best questions and answers from the webinar. Some of these came in before the webinar during registration, some were asked live in the chat, and others were submitted through the Q&A panel.
Licensing and costs
Do I need a separate license for the semantic index? The semantic index is tenant-wide. Once you have a single $30/user/month Microsoft 365 Copilot license active in your tenant, the semantic index gets turned on and content starts getting indexed. Pay-as-you-go alone does not activate it.
What’s the cost for each agent type? There’s no separate cost per agent type. If you have the Microsoft 365 Copilot license ($30/user/month, or $21 for organizations under 2,000 employees), you can create agents using any of the four options. Copilot Studio access is included with that license. End users need to be licensed too, either through the full license, credit packs ($200/month for 25,000 credits), or metered pay-as-you-go.
Is pay-as-you-go a good cost-effective option for most users while one developer holds the full license? It depends on your organization. If you have one $30/user/month license (to activate the semantic index) and everyone else is on metered pay-as-you-go, that can work. But if usage is heavy, you might spend more on metered consumption than you would on per-user licenses. Run the numbers for your situation.
How do I get a dev tenant for testing? The cheapest approach I recommend: get a Microsoft 365 Business Essentials license ($6/month on annual commitment) plus a Copilot license ($21/month for small tenants). That’s about $415/year for a full development environment.
Architecture and capabilities
How secure is it? Can an agent access other people’s data? Everything is security trimmed, just like SharePoint search. Copilot always operates on behalf of the current user. If you don’t have access to a document or mailbox, it won’t show up in results, period.
Can you elevate permissions for an agent? Not within the standard Copilot extensibility model. Copilot is set up to always work on behalf of the current user. There are emerging capabilities around Entra ID agent IDs and digital IDs, but those are still treated like user identities. If you need elevated access, build a REST API with your own auth and expose it as a custom action.
Can you connect to Azure AI Search? Yes. Copilot Studio has a native connector for Azure AI Search. With declarative agents, you’d expose it as a REST API action or connect via an MCP server.
Does searching the web through an agent allow downloading content? Web search through Copilot agents uses Bing to find and reference public content. It doesn’t download files to your environment. It retrieves and references content for the model to use in generating responses.
Agent types and differences
Is Agent Builder the same as declarative agents? No. Agent Builder (formerly Copilot Studio Light) is a simplified, browser-based agent creation experience inside Copilot chat. Declarative agents are built using the Microsoft 365 Agents Toolkit in VS Code. They use different orchestrators and have different capability sets.
Do SharePoint agents work with SharePoint pages? No, only documents in document libraries. Lists may be coming (or may already be available, as this changes frequently), but pages are not currently supported.
How do I deal with ALM (application lifecycle management) for declarative agents? Since declarative agents are just files (JSON, text, YAML), they fit naturally into standard developer workflows: source control, branching, pull requests, CI/CD pipelines. This is one of the strengths of the ATK approach compared to Copilot Studio, where the ALM story is much more limited.
Knowledge and data
Where does embedded knowledge in declarative agents live, and can access be restricted? When you embed knowledge files in a declarative agent using the Agents Toolkit, those files are packaged with the agent and deployed as part of the Teams app. Access is controlled through the app’s deployment and permissions model in your tenant.
Can you create groups of data sources or use metadata to organize knowledge? Not directly as “groups.” The way you manage this is through good naming and descriptions on your connectors and knowledge sources, plus coaching the orchestrator in your agent’s instructions on when to use which source. For declarative agents, you can also add model-only descriptions that are visible to the orchestrator but not the end user.
What about the announcement that Copilot Studio agents will get the same answer quality as Agent Builder? That’s referencing upcoming changes in the 2026 Wave 1 release plan. It means Copilot Studio agents targeting M365 Copilot users will get access to similar orchestration quality. The details of what this changes in practice are still emerging.
Governance and publishing
Can you restrict who creates agents? Yes, there are ways to lock down your environment so not everyone can publish agents. The specific controls depend on which tool you’re using and your tenant admin settings.
Can agents be deployed to unlicensed users? No. Users need to be licensed to use Copilot, whether through the full $30/user/month license, credit packs, or pay-as-you-go consumption.
What about GDPR and privacy? Nothing extra specific to custom agents. GDPR and privacy are handled at the Microsoft 365 platform level. Of course, if your agent collects PII from users or pulls data from external systems, you need to factor that into your data handling practices.
Want to learn more about building declarative agents?
I’ve been training developers on Microsoft 365 Copilot extensibility since the platform launched, and declarative agents are where I see the most opportunity for developer teams right now. If this recap sparked your interest, I’ve got a free webinar coming up next week, April 15, where I’ll go deeper into building declarative agents for Microsoft 365 Copilot using the Agents Toolkit in VS Code.
I’ll walk through demos, explain the architecture in more detail, and answer your questions live. Register for the free webinar here:
Discover the pro-developer path to extending Microsoft 365 Copilot with declarative agents — built in VS Code with the Microsoft 365 Agents Toolkit.
https://www.voitanos.io/webinars/microsoft-365-copilot-developers-guide-declarative-agents-20260415/
When you register, there’s a question asking what you want to learn. Answer it, because I’ll use your responses to shape the content, just like I did for this webinar.
Hope to see you there!
Microsoft MVP, Full-Stack Developer & Chief Course Artisan - Voitanos LLC.
Andrew Connell is a full stack developer who focuses on Microsoft Azure & Microsoft 365. He’s a 22-year recipient of Microsoft’s MVP award and has helped thousands of developers through the various courses he’s authored & taught. Whether it’s an introduction to the entire ecosystem, or a deep dive into a specific software, his resources, tools, and support help web developers become experts in the Microsoft 365 ecosystem, so they can become irreplaceable in their organization.
