Here is the situation. I have spent real time building custom agents. The kind where you contain the agent to a tight set of questions, define the parameters of how it should answer, and lock down the data sources. That work has been valuable. But once the underlying SharePoint structure was sound, we often ended up using the out-of-the-box Copilot agent anyway. The custom version became overkill.

That experience already made me curious. How much of what people are building custom agents to do could be solved by a better-structured site? And now Microsoft has announced Skills, which seem to do exactly that — capture how a team works, in natural language, and reuse it.

So I want to walk through what Microsoft has actually told us about Skills, what custom agents still do that Skills do not, and where I think the open questions sit. I have not had hands-on time with Skills yet. Neither has most of the market. So this is a clear-eyed look at the documentation, not a verdict.

What Skills Are, Based on What Microsoft Has Published

Skills sit inside AI in SharePoint, which is the bigger natural-language experience Microsoft launched in March 2026 and expanded in April. The Skills capability rolled into public preview in April 2026.

According to Microsoft’s documentation, here is what a Skill is:

• A reusable, multi-step workflow that you describe in natural language
• Stored as a markdown file inside the site, in an Agent Assets library
• Created by anyone with edit permission on the site, and run by anyone with view permission
• Loaded automatically when the AI thinks it matches your prompt, or invoked by name
• Limited to actions the user already has permission to do on that site

Microsoft’s published examples include generating quarterly reports from data in the site, drafting proposals using past content, and creating standardised project tracker lists. The framing is consistent: a Skill captures how the team already works and makes it reusable.

What Microsoft Has Explicitly Said Skills Cannot Do

This is the part that matters when comparing to custom agents. Microsoft Learn states it directly: Skills cannot connect to external systems or execute custom code. A Skill can only do what AI in SharePoint already supports on the site, and only what the user running it has permission to do.

That single sentence is the cleanest line we have between Skills and custom agents right now.

What Custom Agents Still Do That Skills Do Not

Custom agents — whether built directly in SharePoint or in Copilot Studio — sit in a different territory. Here is what is documented, not speculative:

• Custom SharePoint agents can use up to 20 source items. Those items can be sites, libraries, folders, or files, and they can come from more than one site. A Skill is scoped to the site it lives in.
• Copilot Studio agents can use sources beyond SharePoint. That includes Dataverse, third-party connectors, and other systems Microsoft has documented support for.
• Custom agents have their own identity. You name them, define their purpose, write their starter prompts, and shape their behaviour. Skills are tools the AI loads; they are not entities you talk to.

So in plain English: if the work stays inside one SharePoint site, Skills are clearly in scope. The further the work travels — across sites, across systems, across business processes — the more a custom agent earns its place.

Where the Honest Open Questions Sit

This is where I want to be careful. The documentation tells us what Skills can do and what they cannot do. It does not tell us how well they do the things they can do, or how reliable they are at the edges. These are the questions I want to test before drawing harder lines:

1. How well does a Skill hold its scope? A custom agent’s value, in my experience, has been containing it tightly. A Skill is loaded by the AI based on prompt matching, which is a softer kind of containment. I want to see how that behaves in practice.
2. How well do Skills handle ambiguity? Microsoft’s own documentation notes that the Skill’s quality depends on how it is written. Vague Skills will produce vague outputs. That is a real consideration before assuming Skills replace anything.
3. How does the markdown-in-the-site model scale? Skills live in an Agent Assets library on the site. That is excellent for governance and version control. It also means Skills proliferate per site, which has its own management story to think about.
4. How does this evolve? Microsoft has said custom AI Skills, a separate and broader concept that organisations can define, are coming. The boundary between site-level Skills and tenant-level Skills is not fully drawn yet.

So, Do Skills Replace Custom Agents?

Based only on what is documented today, I would say this:

• If you are about to build a custom agent whose only job is to answer questions about one site, in a particular way, using that site’s content — pause. That is exactly the territory Skills are designed for.
• If you are about to build a custom agent that needs to span multiple sites, pull from non-SharePoint sources, or trigger work outside SharePoint — Skills are not a replacement. Microsoft’s documentation rules that out explicitly.

The honest answer is that the line between them is becoming clearer, but the testing has not caught up to the announcements yet. Anyone telling you they know exactly where Skills replace custom agents and where they do not is moving faster than the evidence currently supports.

What I Would Do Right Now

If your organisation is partway through a custom agent build, I would not abandon it on the basis of Skills being announced. The preview is still rolling out. Real-world testing is still thin.

What I would do is this. For any new agent that goes into the build queue, ask whether the work it does crosses outside one SharePoint site. If it does not — and many do not — then Skills are worth at least a pilot before committing to the cost and timeline of a custom agent. If it does, custom agents remain the right home for that work.

That is a practical position you can take today, with the evidence we currently have.

The Bigger Picture

Microsoft is steadily moving the easy work down the stack — into the platform, into natural language, into things that do not need a developer or a Copilot Studio licence. That is the pattern Skills fit into. It is also the pattern behind the new SharePoint experience, AI in SharePoint, the SharePoint Admin Agent, and the broader push to make Copilot useful without bespoke builds.

What that means for those of us watching it: the case for custom development gets narrower over time, but the case for clean content structure gets stronger. Skills, agents, Copilot — all of them work better when the underlying SharePoint is in order. That part of the story has not changed and is not going to.