Add post: Modern Agents Have Skills Now in Copilot Studio by Roelzz · Pull Request #329 · microsoft/mcscatblog

Roel Schenk (Roelzz) · 2026-06-15T18:50:45Z

Summary

New post introducing Skills in modern Copilot Studio agents: an on-demand procedural layer based on the Agent Skills open format.
Covers why builders should care (manageability, context window, sometimes accuracy), the add/invoke UX, how Copilot Studio Skills differ from coding-agent Skills, how Skills tighten the agent loop and direct tool use, and when to use Skills vs instructions.
Two Mermaid diagrams (on-demand selection, Skills-vs-instructions decision path) and three screenshots.

Checklist

Ran /review-post and addressed feedback
Local server renders correctly (./tools/run.sh)
All images have alt text and captions
2-3 internal links to related posts
Tags chosen for Chirpy "Further Reading" overlap

…tudioNow Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

github-actions · 2026-06-15T18:51:57Z

Blog preview check passed for 08a8f5e.

The Jekyll site compiled successfully. No newly added _posts/*.md files were detected in this PR.

adilei

Thanks!

adilei · 2026-06-15T20:26:35Z

+mermaid: true
+---
+
+Enterprise agents rarely fail for lack of facts. They fail because the organization has a specific way of doing the work, and that procedure ends up scattered across one ever-growing system prompt. Modern agents have a cleaner place to put it: **Skills**.


It's a nice intro but "rarely fail" is a little clickbaity and somewhat untrue :)

adilei · 2026-06-15T20:29:03Z

+
+Enterprise agents rarely fail for lack of facts. They fail because the organization has a specific way of doing the work, and that procedure ends up scattered across one ever-growing system prompt. Modern agents have a cleaner place to put it: **Skills**.
+
+If you have spent time with coding agents recently, you have probably already met them. A Skill is a small, self-contained package of procedural knowledge: a `SKILL.md` file with a name, a description, and instructions that tell the agent how to do a specific kind of task. The agent loads it only when a task actually needs it.


I think this is too soon to introduce "procedural knowledge". It's a loaded term that should be explained, probably later?

adilei · 2026-06-15T20:30:13Z

+
+Skills are based on the [Agent Skills open format](https://agentskills.io/), an open standard originally developed by Anthropic. The shape is deliberately simple, but the benefits are real:
+
+- **Manageability.** Instead of one ever-growing instruction blob, each procedure lives in its own focused Skill. You can reason about, review, and version them one at a time.


Again, procedure. Better to frame skills as instructions (which is what they are) and talk about procedural knowledge in the section that dicusses when to use skills

adilei · 2026-06-15T20:31:42Z

+
+- **Manageability.** Instead of one ever-growing instruction blob, each procedure lives in its own focused Skill. You can reason about, review, and version them one at a time.
+- **You avoid saturating the context window.** Skills load *on demand*. The agent keeps only the names and descriptions in view by default, and pulls the full instructions into context only when a task matches. Ten procedures cost you ten short descriptions, not ten full playbooks, in every turn.
+- **Sometimes accuracy.** This one is genuinely use-case dependent. In specific scenarios, like large toolsets or sophisticated retrieval, keeping irrelevant procedure out of context and bringing in the right one can improve how reliably the agent acts. It is not a guarantee. Treat it as something to evaluate, not assume.


this isn't clear: n specific scenarios, like large toolsets or sophisticated retrieval. I know this is my comment verbatim, but it the sentence ends up being awkward here

adilei · 2026-06-15T20:32:46Z

+
+## The same benefits show up in Copilot Studio
+
+The good news is that the modern Copilot Studio orchestrator works the same way: it can reason over a set of available Skills, select the relevant one, and bring its procedure into context for that task. So the manageability and context benefits carry over directly, and the accuracy question is still yours to evaluate.


"that task" -- which task?

adilei · 2026-06-15T20:51:08Z

+
+## What Skills look like in practice
+
+Skills shine wherever the *right answer depends on following a sequence*. A few shapes worth recognizing:


I think this is the wrong framing. I think what you want to say is: here are some useful patterns that leveraging skill. The whole bit about following a sequence/being a procedure is a little reparative and not always informative. E.g. is generating a file based on a template a procedure? A bit of a strech.

adilei · 2026-06-15T20:52:00Z

+- **Generating a file from a template.** The output must follow a fixed structure every time, like a report, a summary, or a standardized record. The Skill carries the template and the rules for filling it in.
+- **A specific order of execution.** The task is a pipeline: discover available processes, then run the analysis, then add the ROI pre-scan, then format the result. The Skill keeps the steps in the correct sequence and points at the right tool at each one, exactly what the reasoning view above is showing.
+
+In every case the Skill is the *playbook*, not the data and not the system access. Knowledge gives the agent facts, tools give it reach, and Skills give it the operating procedure that ties them together.


System access?

adilei · 2026-06-15T20:53:03Z

+
+## A Skill, or a new agent?
+
+Before Skills, the instinct for every new process was to build another specialized agent: one for HR leave, one for IT incidents, one for refunds. Sometimes that is right: security boundaries, distinct audiences, and clear business ownership still justify a separate agent.


But the examples you give here are ones that do justify seperate agents.

adilei · 2026-06-15T20:55:32Z

+
+Before Skills, the instinct for every new process was to build another specialized agent: one for HR leave, one for IT incidents, one for refunds. Sometimes that is right: security boundaries, distinct audiences, and clear business ownership still justify a separate agent.
+
+But often the real need is not a new agent, it is a new *procedure*. If the same agent serves the same audience, shares the same knowledge boundary, and already has the right tools, a Skill is the better unit of modularity. You are not building another agent to maintain; you are teaching the existing one another way of working. That is how Skills cut down on agent sprawl.


I know I've already said that, but I'm really not sure "procedure" is doing the work we want it to do for the reader. Is driving the conversation in a certain way (e.g. questions to ask) a procedure? Procedure sounds too much like policy or a playbook, which gives the wrong connotation I think.

It is useful to explain "common sense" vs "how the organization does things" but I think there's an inflation of this term here

adilei · 2026-06-15T20:56:17Z

+
+## A word on trust
+
+Because a Skill shapes how the agent behaves, it is a trust surface, especially one copied from a community source, generated by AI, or reused from another environment. Review Skills before adding them: keep them focused (a Skill is not a place to dump an entire policy manual), avoid overlapping vague Skills that confuse routing, and check for prompt injection or instructions to misuse tools.


Couple of different points get conflated here. If trust is important, stay on topic

…dback - Replace 'procedural memory' framing with 'instructions and resources loaded on demand' - Add mental-model analogy section (table + hub diagram) and colored context-flow diagrams - Update for shipped support: bundled resources and executable scripts - Address review feedback (intro, accuracy/speed bullets, tool soft-pointing, agent-vs-skill examples, trust, dead link) - New header image Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

adilei · 2026-06-20T09:37:06Z

Hi Roel Schenk (@Roelzz) — adilei did an editor's pass on this post and I've opened it as a separate branch on the main repo so you can review and approve before we publish: #332

It builds on your PR with some tightening (intro reframe, dropped the standalone Speed benefit, reworked the 'Skill vs. new agent' heuristic), a few accuracy fixes, and the required agent_edition: modern front matter. Full change list is in #332.

If you're happy with the edits, approve #332 and we'll merge it. If you'd rather keep iterating here on your own branch, just say the word and we'll port anything useful back over.

adilei · 2026-08-05T12:31:18Z

Closing this one in favor of the other Skills post already in review (see #321 / #328) so we don't duplicate coverage — let's consolidate there. Thanks Roel Schenk (@Roelzz)!

Add post: Modern Agents Have Skills Here's How They Work in Copilot S…

56a5ca3

…tudioNow Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

adilei reviewed Jun 15, 2026

View reviewed changes

adilei mentioned this pull request Jun 20, 2026

Review edits: Modern Agents Have Skills Now (for @Roelzz to approve) #332

Merged

adilei closed this Aug 5, 2026


		Enterprise agents rarely fail for lack of facts. They fail because the organization has a specific way of doing the work, and that procedure ends up scattered across one ever-growing system prompt. Modern agents have a cleaner place to put it: Skills.

		If you have spent time with coding agents recently, you have probably already met them. A Skill is a small, self-contained package of procedural knowledge: a `SKILL.md` file with a name, a description, and instructions that tell the agent how to do a specific kind of task. The agent loads it only when a task actually needs it.


		Skills are based on the [Agent Skills open format](https://agentskills.io/), an open standard originally developed by Anthropic. The shape is deliberately simple, but the benefits are real:

		- Manageability. Instead of one ever-growing instruction blob, each procedure lives in its own focused Skill. You can reason about, review, and version them one at a time.


		## The same benefits show up in Copilot Studio

		The good news is that the modern Copilot Studio orchestrator works the same way: it can reason over a set of available Skills, select the relevant one, and bring its procedure into context for that task. So the manageability and context benefits carry over directly, and the accuracy question is still yours to evaluate.


		## What Skills look like in practice

		Skills shine wherever the right answer depends on following a sequence. A few shapes worth recognizing:


		## A Skill, or a new agent?

		Before Skills, the instinct for every new process was to build another specialized agent: one for HR leave, one for IT incidents, one for refunds. Sometimes that is right: security boundaries, distinct audiences, and clear business ownership still justify a separate agent.


		Before Skills, the instinct for every new process was to build another specialized agent: one for HR leave, one for IT incidents, one for refunds. Sometimes that is right: security boundaries, distinct audiences, and clear business ownership still justify a separate agent.

		But often the real need is not a new agent, it is a new procedure. If the same agent serves the same audience, shares the same knowledge boundary, and already has the right tools, a Skill is the better unit of modularity. You are not building another agent to maintain; you are teaching the existing one another way of working. That is how Skills cut down on agent sprawl.


		## A word on trust

		Because a Skill shapes how the agent behaves, it is a trust surface, especially one copied from a community source, generated by AI, or reused from another environment. Review Skills before adding them: keep them focused (a Skill is not a place to dump an entire policy manual), avoid overlapping vague Skills that confuse routing, and check for prompt injection or instructions to misuse tools.

Uh oh!

Conversation

Roel Schenk (Roelzz) commented Jun 15, 2026

Summary

Checklist

Uh oh!

github-actions Bot commented Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

adilei left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

adilei commented Jun 20, 2026

Uh oh!

adilei commented Aug 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

github-actions Bot commented Jun 15, 2026 •

edited

Loading