Craig Stanley
Home / Capabilities / Agent Builder / Agent Builder limits worth knowing before you start

Agent Builder limits worth knowing before you start

The knowledge source limits and known gaps in Agent Builder, and how they shape what kind of agent you should build there.

· 3 min read · Craig Stanley
In short, explained

Agent Builder helpers can only read so much. Knowing the limits first stops you building something that can't hold all the papers it needs.

Agent Builder agents can use up to four websites, 100 SharePoint files, one SharePoint list, 50 OneDrive files, five Teams chats and five meetings, and 20 uploaded files. They can't be used in Teams group or one-to-one chats. If your idea needs more, plan for Copilot Studio.

Per-agent knowledge limits: 4 public URLs (≤2 path levels, no query strings), 100 SharePoint files, 1 list (20,000 rows / 50 MB raw text, truncated beyond), 50 OneDrive files, 5 chats + 5 meetings, 20 embedded files (512 MB; 30 MB Excel). No CMK, no Teams group/1:1 chat use; Restricted SharePoint Search blocks SharePoint knowledge.

The numbers

Microsoft's Learn page on knowledge sources in Agent Builder gives these limits for each agent.

Knowledge sourceLimit per agent
Public website URLs4
SharePoint files, folders or sites100 files
SharePoint lists1 list
OneDrive files50 files
Teams chats and meetings5 chats and 5 meetings
Files uploaded from your device20 files
Copilot connectorsNo documented number; an admin must enable them

A few details sit underneath those numbers. Public website URLs can only go two levels deep and can't contain query parameters, so a page like example.org/a/b/c won't be accepted. A SharePoint list can hold up to 20,000 rows and 50 MB of raw text; beyond that, Microsoft says the agent truncates it and says so in its answer. Choosing a SharePoint site doesn't include its lists, and list attachments aren't used. Uploaded Word, PDF and PowerPoint files can be up to 512 MB, and Excel files up to 30 MB.

Known gaps

Microsoft lists some limitations in the Agent Builder overview. Agents built there can't be used in Teams group chats or one-to-one chats. Customer-managed keys aren't supported. If Restricted SharePoint Search is turned on, SharePoint can't be used as a knowledge source. And if an admin has blocked web search in Copilot, web content is blocked as knowledge even though the toggle still looks available.

There are also file behaviours that fail quietly. Microsoft's table says a file protected with Double Key Encryption is embedded but not used as knowledge, with no error shown. Files with user-defined permissions or with extract rights disabled can make agent creation fail without an error message.

How the limits shape the design

I read the limits as a description of the right kind of agent for Agent Builder: one job, a small and well-chosen set of sources, used by people in the Copilot app. An agent that needs to answer from a whole document library, live system data or a big list of websites is probably a Copilot Studio agent from the start.

The limits also help with a quieter problem. An agent with 100 loosely related files gives vaguer answers than one with ten good ones. Microsoft's own guidance is to keep the contents of files you provide concise. So a limit can push you towards a better agent.

A worked example

This scenario is illustrative. An HR team wants an agent that answers leave questions. Their plan:

SourceCountWithin limits?
Leave policy and three related policies on SharePoint4 filesYes
The bank holiday page on gov.uk1 URLYes, if it's two levels or fewer
A list of 1,400 team calendars1 listYes, under 20,000 rows
Every HR document in the libraryAbout 600 filesNo, over 100

They drop the last line and keep the agent to leave questions. If they later want one agent for every HR question, that's the point to look at Copilot Studio, where knowledge limits are different.

What I'm still checking

The table shows what can be added. I haven't found Microsoft guidance on how answer quality changes as an agent approaches 100 files. I'd test it with real questions before assuming more sources means better answers.

Sources

Read next

A question to take awayWhich of these do you already pay for and not use?

About me

Craig Stanley

Microsoft AI consultant and technical architect, based in Whitley Bay. Over the last few years I've delivered Microsoft 365 Copilot, Copilot Studio agents, Microsoft Foundry (formerly Azure AI Foundry) work and governance for UK public sector and financial services organisations.

What interests me is the decision underneath the tool: what it costs, what it risks, and whether a small, transparent model can make it better. I write the methods up here and on Substack so anyone can use them.

I write this site to learn in public: explaining each idea simply is how I check I understand it. Why I write this site.

Find me