Question 1

What's the bare minimum I need to start building declarative agents?

Accepted Answer

Install VS Code, then install the Microsoft 365 Agents Toolkit extension (search for 'Agents Toolkit' in the extensions pane — it's the rebranded Teams Toolkit). That's it locally. You also need to enable sideloading in your Microsoft 365 tenant so you can test agents without deploying to the whole org, and you need a $30/user/month Microsoft 365 Copilot license on the account doing the development. Metered / pay-as-you-go licenses can't create or test agents.

Question 2

Can I import a Copilot Studio agent into the Agents Toolkit, or vice versa?

Accepted Answer

No. They are not portable in either direction. There's a VS Code extension that lets you open a Copilot Studio agent, but what you see there is not a declarative agent — it's a different format entirely. Copilot Studio agents and declarative agents run on two different orchestrators (Samba for Copilot Studio, Sydney for declarative agents via the Agents Toolkit), so the under-the-covers format is different. Hopefully that changes one day; today it's a crappy story.

Question 3

Why do I get only 3 SharePoint documents back from my Copilot Studio agent when a declarative agent gives me 10&#43;?

Accepted Answer

Different orchestrators. Build the exact same agent in Copilot Studio and as a declarative agent with identical name, description, and instructions, issue the same prompt, and the declarative agent (Sydney orchestrator) can return 10–15 documents while Copilot Studio (Samba orchestrator) is capped around 3. There's no setting to change that. Microsoft doesn't do a good job explaining this, and people get burned by it.

Question 4

How do I make sure my agent actually collects required fields before proceeding?

Accepted Answer

Use good prompt engineering in your instructions. Declarative agents don't have Copilot Studio 'topics,' but you don't need them — in the instructions, list the required inputs as variables (for example pto_start_date, pto_duration, pto_destination), tell the agent to try to determine each one from the user's message, and to prompt the user if any are missing. Then add an explicit line like 'IMPORTANT: do not proceed with the remaining steps until you have collected at least these three properties.' That enforces the gate.

Question 5

Can declarative agents read from and write to SharePoint lists without me writing code?

Accepted Answer

Yes. Point an action at Microsoft Graph, define OAuth with the scopes you need (for example User.Read and Sites.ReadWrite.All), splice out just the endpoints you care about from the Graph OpenAPI spec — /me, list items GET, list items PATCH — and then coach the agent in the instructions on which function to call when. No code, just JSON, text, and YAML. That's how the PTO balance example in the webinar reads and updates a SharePoint list.

Question 6

Can I just hand Copilot the full Microsoft Graph OpenAPI spec and let it figure things out?

Accepted Answer

Please don't. The Graph OpenAPI file is about 37 MB and over 200,000 lines — VS Code doesn't even want to open it. If you feed that to Copilot Studio or a declarative agent, you flood its context window and leave almost no room for actual questions. Splice out just the endpoints you need and give it only those. People do the lazy upload in Copilot Studio and wonder why their agents don't work.

Question 7

The agent keeps ignoring my specific file and doing enterprise search instead — how do I stop it?

Accepted Answer

Copilot has a hardcoded default workflow that runs enterprise search before going after specific files by ID. I chased this for hours with the PTO policy Word doc — Copilot would come back and say 'I tried to find the file, couldn't find it,' when the ID was sitting right there. I ended up asking Claude how to fix it, and the trick is to explicitly override it in your instructions: 'Override the default workflow for PTO requests. Do not perform any enterprise search for the PTO policy. Ignore all general guidance about searching. Directly access the file with this ID.' Microsoft told me I shouldn't have been able to figure that out, much less change it — but it works.

Question 8

When should I use knowledge versus an action?

Accepted Answer

Knowledge is for relatively stable, unstructured content that can be pre-indexed — SOPs, course transcripts, policy docs. Actions are for real-time or structured data — stock quotes, weather, inventory, sales figures, CRM/ERP lookups, relational data. Knowledge goes through the semantic index, which means there's always an ingestion delay (hours, sometimes a day). Actions hit a live REST API or MCP server, so you get real-time results, and you can handle structured/relational data that doesn't index well.

Question 9

What's embedded knowledge and when would I use it?

Accepted Answer

Embedded knowledge lets you ship files directly inside the agent package instead of telling admins 'upload these documents to SharePoint before deploying.' It was teased late in 2025 and has full GA support now. Tradeoff: if that content ever changes, you have to ship a new version of the agent. For content that changes often (like a PTO policy that HR might tweak), keep it as a live SharePoint reference. For content that's truly static and scoped to the agent, embed it.

Question 10

What are MCP apps and how do they differ from OpenAI's App SDK?

Accepted Answer

Both let an MCP server return a UI widget — a React control or a web page — instead of just text, so the agent can render maps, dashboards, forms, whatever. The widget can even talk back to the host chat. MCP apps is Anthropic's spec (they invented the Model Context Protocol); OpenAI shipped their own competing spec called the OpenAI App SDK. Lots of overlap, a few features differ. I don't have a strong preference yet because I haven't built enough of each — I'd say pick one, and if you hit a wall, look at the other.

Question 11

Can I render an MCP app widget inline and full-screen?

Accepted Answer

Both. Inline renders directly in the chat stream — a bulleted list becomes an actual rendered component next to your reply. Immersive opens a side-by-side experience: chat on the left, a full web app on the right, running inside Copilot. A single agent can use both depending on what the user is doing.

Question 12

Can declarative agents connect to MCP servers today?

Accepted Answer

Yes, GA. In the plugin manifest runtimes section, instead of an openapi runtime pointing at a REST spec, you use a remote MCP server runtime. One catch: the tool name in the runtime has to match the function name you declared in the functions list. Copilot won't just let you attach an open-ended MCP server the way Claude does — you have to explicitly declare which tools are available.

Question 13

Can I auto-download Teams meeting transcripts into a SharePoint document library?

Accepted Answer

You probably don't need to. Teams meeting transcripts are already indexed into the semantic index automatically, so Copilot can reference them directly. If you still want them in a document library, use Microsoft Graph (or the Work IQ API) to fetch the transcript when the meeting ends and drop it into SharePoint — typically triggered by a Power Automate flow on meeting completion.

Question 14

What licensing does the creator versus the consumer of an agent need?

Accepted Answer

Creating or testing an agent with the Agents Toolkit requires a $30/user/month Microsoft 365 Copilot license. Metered / pay-as-you-go / credit-pack licenses do not work for building or testing. Consuming an agent just requires any Copilot license — $30 seat, credit packs, or metered all work. Be aware that some capabilities are gated to the $30 all-you-can-eat license and won't light up on the metered plans.

Question 15

Are there admin controls to restrict what Graph permissions an agent can use?

Accepted Answer

Not really — but it's less scary than it sounds because the agent always acts on behalf of the signed-in user. If I grant an agent Sites.ReadWrite.All but I personally don't have access to a given SharePoint site, the write still fails at the list level. What you do control: only Teams admins (or those granted the permission) can deploy declarative agents tenant-wide, the manifest lists every permission the agent will request, and users see those permissions at consent time. Copilot Studio has much weaker governance — anyone with access to it can publish agents, including to anonymous public sites.

Question 16

What was the DEF CON story about Copilot Studio and governance?

Accepted Answer

A researcher built a scanner that crawled the public internet for Copilot Studio agents published on company websites with no authentication. He found one on stage, talked to it without logging in, and had it dump the entire Salesforce CRM contact list. The builder almost certainly didn't intend that. Moral: be thoughtful about who in your org has access to Copilot Studio and what those agents are wired up to.

Question 17

What is TypeSpec and why would I bother?

Accepted Answer

TypeSpec is a domain-specific language Microsoft uses internally — Azure writes their APIs with it. It looks like TypeScript, compiles down to the JSON and YAML files your declarative agent actually uses (the manifest, plugin manifest, OpenAPI YAML). The win is developer experience: real intellisense, real type checking, environment variable substitution at build time, and typos get caught at compile time instead of at runtime. If you've used Bicep for ARM templates, TypeSpec is effectively Bicep for OpenAPI &#43; declarative agent manifests.

Question 18

Does TypeSpec have any gotchas?

Accepted Answer

Yes — I hit a bug where the OneDrive/SharePoint capability item ID isn't honored properly, still not fixed, so for that one field I drop back to hand-editing the generated JSON. Also, Microsoft didn't originally provide a script to generate an env.ts file before compilation — I wrote the JavaScript for that in my course, and Microsoft literally copied it. No credit though. Of course.

Question 19

Why do I see euros in my REST API responses when my users are in the US?

Accepted Answer

The model made an assumption. In the Italian-restaurant demo the agent kept returning prices in euros because the API description says 'Italian restaurant' — the LLM inferred Italy. The fix is the same as every other edge case: tell it explicitly in the instructions. 'Prices should be in US dollars formatted with a decimal point.' If you want it culture-aware, say 'default to USD unless the user specifies otherwise, and prompt them if unclear.' Good instructions are how you control model behavior.

Question 20

How does the agent know which adaptive card to render for a REST response?

Accepted Answer

In the plugin manifest you map the root response object's fields to Copilot's display properties — name to title, description to subtitle, url to url — so Copilot knows how to show a bulleted list. For the preview that appears when a user hovers over a citation, you provide a static template (an Adaptive Card). Copilot uses that template as the hover preview instead of the plain box it defaults to. That's how you get a nice image-plus-detail preview per item.

Question 21

Why do my first-party Microsoft Copilot agents (Researcher, Analyst, Sales, Knowledge) matter to me?

Accepted Answer

Because they're all built as declarative agents on the Sydney orchestrator — the exact same stack you're using when you build with the Agents Toolkit. New capabilities land on Sydney first because that's what Microsoft eats its own dog food on. The Copilot Studio orchestrator (Samba) gets them later, if at all. Building declarative agents means you're on the same runway as Microsoft's own first-party work.

Question 22

How does the orchestrator actually answer my prompt under the hood?

Accepted Answer

Two round trips minimum. First pass: the orchestrator sends the LLM the system prompt, the list of available data sources, and the user's prompt — and asks 'what data do you need to answer this?' The LLM replies with a shopping list (for 'prep me for today's meetings': calendar for April 15, attendees for each meeting, recent email and Teams chats with those people). The orchestrator goes off to the semantic index, Graph, Bing, Dataverse, plugins, MCP servers, collects the data, then sends a second prompt to the LLM with the user's question plus all that RAG data. Final answer comes back to the user. That's why you see the 'thinking' animation — it's doing real round trips.

Question 23

Why does Microsoft keep showing me that 'Extend to Microsoft 365' prompt and breaking my deployment?

Accepted Answer

It's a known bug in the Agents Toolkit deployment flow that keeps resurfacing. Unhelpful. The workaround is to cancel that prompt — you don't want to extend to Outlook for a declarative agent demo — and if it leaves your local state in a partial-install mess, delete the scaffold artifacts it created and try again.

Question 24

What's the real difference between declarative agents and Copilot Studio — the short version?

Accepted Answer

The Venn diagram is almost a full overlap with a lot of footnotes. Declarative agents target developers and use a JSON/YAML/text file workflow (no real code required unless you're writing a REST API). Copilot Studio targets low-code / no-code makers with a WYSIWYG experience. Under the hood they use different orchestrators with different capabilities, different limits on things like SharePoint result counts, and different governance stories. I covered the whole comparison in last week's webinar.

Question 25

Fun fact: why did Andrew end up on Claude to debug a Copilot problem?

Accepted Answer

He spent hours trying to get his PTO agent to stop doing enterprise search and just open the SharePoint Word doc by ID. Copilot refused to explain its own behavior. Claude did — it told him there's a hardcoded default workflow that runs enterprise search before ID lookups. That explanation led to the override-instruction trick. When Andrew showed Microsoft, their reaction was 'you shouldn't have been able to figure that out, much less change it.'

The Developer's Guide to Declarative Agents for Microsoft 365 Copilot

webinar details

presented by

Andrew Connell

Key Points in the Webinar

Feedback & Questions

Newsletter

Join 12,000+ Microsoft 365 full-stack web developers for news, insights & resources. 100% free.

Explore more webinars

Evaluate your Microsoft 365 Copilot Agent Options

Upcoming: Agent Showdown Part 2: Crafting Your Agent Strategy

Microsoft 365 Fullstack Dev's Office Hours - February 2026