Add Copilot Agent Advisor skill

[SandraBcna (SandraBcna)]

Add Copilot Agent Advisor skill

What this adds

A new skill, copilot-agent-advisor, that recommends which Microsoft 365 / Copilot Studio agent option fits a described scenario:

• Use Microsoft 365 Copilot as-is (no build)
• Standard agent (declarative agent — Copilot's own orchestrator + models, inside M365)
• Custom agent (custom engine agent — bring your own orchestrator/models, custom workflows, own hosting, external reach)

It also advises which Copilot Studio harness to build on — GitHub Copilot, standard, or Copilot chat — and flags that agents aren't transferable between the GitHub Copilot and standard harnesses.

Scope and differentiation

This skill focuses on the upstream decision — whether to build at all, and which of the three agent types (M365 Copilot as-is, declarative agent, or custom engine agent) — advising the Copilot Studio harness at a decision level rather than going deep on credit/licensing math. It's authored platform-neutral to run in Cowork, Copilot Studio, and Scout.

Why

Makers repeatedly ask "declarative or custom engine?" and "which harness?" This skill turns the decision into a short, factor-based recommendation (knowledge location, models, orchestration, channels, autonomy, collaboration, compliance, speed/skill) with a runner-up, a how-to-build path, and the cost/hosting/compliance trade-offs.

Grounding

Decision logic is grounded in current Microsoft Learn guidance and re-verified against it (per adilei's review):

• "Agents for Microsoft 365 Copilot" — the declarative vs custom engine build approaches.
• Copilot Studio harnesses overview — GitHub Copilot, standard, and Copilot chat harnesses (this replaced the earlier "classic vs new experience" framing).
  The skill also carries a guardrail to re-verify fast-moving specifics (previews, pricing, Copilot Credits) against Microsoft Learn before asserting them.

Validation evidence

• Static quality score: 96/100 (Excellent) — all structural checks pass, no safety flags, clears its launch floor.
• Behavioral tests: 4/4 passed — 2 should-trigger, 2 should-not-trigger:
  • Internal, M365-grounded memo-drafting scenario (drafts monthly product-review memos from the previous memo, meeting transcripts, and dashboard exports; used in Teams; human-in-the-loop; external integrations and autonomy out of MVP scope) → correctly recommended a standard/declarative knowledge agent in M365 Copilot Agent Builder, reached independently from the raw scenario and correctly deferring the out-of-scope custom-engine triggers.
  • Public-facing loan assistant (own fine-tuned model, external APIs, public web/app, proactive notifications) → correctly recommended a custom engine agent (discriminates, doesn't default to the lighter option).
  • "What can Copilot do in Excel?" → correctly does not trigger (general capability question).
  • "Now go build it for me" → correctly holds its "advise, don't build" boundary.

Submission contents

submissions/copilot-agent-advisor/

• SKILL.md — the skill (single file; no scripts/references, so it ships unpacked)
• README.md — human-facing detail-page overview
• metadata.json — catalog sidecar (platforms: Cowork, Copilot Studio, Scout; tags: copilot-studio, agents, decision-support, architecture, advisor)

Platform coverage

The skill is authored to be platform-neutral — its logic is pure advisory reasoning with no dependency on platform-specific tools (clarifying questions, comparison tables, and optional web lookups are described as capabilities, not named tools), so it runs the same in Cowork, Copilot Studio, and Scout agents.

Checklist

• authorUrl set to my GitHub profile (for correct attribution / skillbot mention)
• CLA signed (Microsoft CLA bot on first PR)
• npm run check:submissions passes locally

[Copilot]

Pull request overview

Adds a new submission to the submissions/ gallery: copilot-agent-advisor, an advisory skill that helps users choose between Microsoft 365 Copilot as-is vs a standard (declarative) agent vs a custom (custom engine) agent, and also recommends Copilot Studio’s new vs classic authoring experience (including the no-migration caveat).

Changes:

• Added the copilot-agent-advisor skill instructions (SKILL.md) with factor-gathering, decision logic, and output/guardrails.
• Added a human-facing overview page (README.md) for the skill’s gallery detail page.
• Added catalog metadata (metadata.json) for platforms, tags, author attribution, and versioning.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
submissions/copilot-agent-advisor/SKILL.md	New agent-facing skill defining the decision interview, recommendation logic, and guardrails.
submissions/copilot-agent-advisor/README.md	Human-facing overview and example for the gallery detail page.
submissions/copilot-agent-advisor/metadata.json	New catalog entry metadata (name/description/platforms/tags/author/version/dates).

[SandraBcna (SandraBcna)]

@microsoft-github-policy-service agree company="Microsoft"

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (3)

submissions/copilot-agent-advisor/SKILL.md:37

• The markdown table rows start with ||, which most markdown renderers interpret as an empty first column (misrendered table). Change these rows to start with a single leading | and ensure each row ends with a trailing | for consistent rendering.

| Option | What it is | Use when |
|--------|-----------|----------|
| **Use Microsoft 365 Copilot as-is** | The built-in Copilot experience over Microsoft Graph data, no build. | Built-in capabilities + your Microsoft 365 data already cover the need; you don't need extra knowledge, custom instructions, or actions. |
| **Standard agent (declarative agent)** | Copilot configured with your instructions, knowledge, and actions, running on **Copilot's own orchestrator and foundation models**, inside Microsoft 365 apps. | You want to tailor Copilot for a focused scenario, stay inside Microsoft 365 (Teams, Outlook, SharePoint, Word), inherit Microsoft 365 compliance/RAI, and ship fast with low-code or a streamlined pro-code path. |
| **Custom agent (custom engine agent)** | A fully custom agent where you **bring your own orchestrator and models**, with custom/complex workflows, its own hosting, and multi-channel reach. | You need custom orchestration, your own/fine-tuned/domain models, deterministic complex business logic, proactive/autonomous behavior, group collaboration, delivery outside Microsoft 365, or you're integrating an existing external bot. |

submissions/copilot-agent-advisor/README.md:11

• Same markdown table issue as in SKILL.md: the || prefix creates an unintended empty column in many renderers. Use a single leading | (and preferably a trailing |) so the table renders as intended.

| Option | What it is |
|--------|-----------|
| **Use Microsoft 365 Copilot as-is** | The built-in Copilot experience over your Microsoft 365 data — no build. |
| **Standard agent (declarative)** | Copilot tailored with your instructions, knowledge, and actions, running on Copilot's own orchestrator and models, inside Microsoft 365 apps. |
| **Custom agent (custom engine)** | A fully custom agent — your own orchestrator and models, custom workflows, its own hosting, and reach beyond Microsoft 365. |

submissions/copilot-agent-advisor/README.md:39

• This sentence reads as a guarantee that the skill will always verify details against Microsoft Learn, but the skill text itself conditions verification on whether the platform has web-browsing/lookup capability. Consider rephrasing to match that constraint (e.g., 'when web lookup is available, verify...; otherwise flag as time-sensitive and point to Microsoft Learn') to avoid overpromising behavior.

- **It stays current.** Agent capabilities and previews move fast, so it verifies fast-moving specifics against Microsoft Learn before asserting them.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (3)

submissions/copilot-agent-advisor/SKILL.md:37

• The table rows start with a double pipe (||), which renders as an extra empty first column in many Markdown parsers. Change these rows to start with a single | so the table renders consistently across platforms.

| Option | What it is | Use when |
|--------|-----------|----------|
| **Use Microsoft 365 Copilot as-is** | The built-in Copilot experience over Microsoft Graph data, no build. | Built-in capabilities + your Microsoft 365 data already cover the need; you don't need extra knowledge, custom instructions, or actions. |
| **Standard agent (declarative agent)** | Copilot configured with your instructions, knowledge, and actions, running on **Copilot's own orchestrator and foundation models**, inside Microsoft 365 apps. | You want to tailor Copilot for a focused scenario, stay inside Microsoft 365 (Teams, Outlook, SharePoint, Word), inherit Microsoft 365 compliance/RAI, and ship fast with low-code or a streamlined pro-code path. |
| **Custom agent (custom engine agent)** | A fully custom agent where you **bring your own orchestrator and models**, with custom/complex workflows, its own hosting, and multi-channel reach. | You need custom orchestration, your own/fine-tuned/domain models, deterministic complex business logic, proactive/autonomous behavior, group collaboration, delivery outside Microsoft 365, or you're integrating an existing external bot. |

submissions/copilot-agent-advisor/README.md:11

• Same table formatting issue as in SKILL.md: using || at the start of each row can introduce an unintended blank column. Use a single leading | per row for reliable rendering.

| Option | What it is |
|--------|-----------|
| **Use Microsoft 365 Copilot as-is** | The built-in Copilot experience over your Microsoft 365 data — no build. |
| **Standard agent (declarative)** | Copilot tailored with your instructions, knowledge, and actions, running on Copilot's own orchestrator and models, inside Microsoft 365 apps. |
| **Custom agent (custom engine)** | A fully custom agent — your own orchestrator and models, custom workflows, its own hosting, and reach beyond Microsoft 365. |

submissions/copilot-agent-advisor/SKILL.md:99

• “standard-harness licensing” reads like an internal term and could be confusing or stale if copied elsewhere. Consider rephrasing to a more user-facing description (e.g., ‘licensed via Copilot Studio licensing’ / ‘licensed separately from Copilot Credits’) and keep the definitive details in the linked Microsoft Learn reference.

- **Standard harness — dependable and rule-based.** You define the topics, prompts, and paths so a structured, repeatable conversation or workflow responds **predictably**, drawing on your prompt library and enterprise knowledge. Billed via **standard-harness licensing**. Recommend it when the scenario is well-defined and you want consistent answers (e.g. an internal help-desk agent that answers common questions and routes simple requests).

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (3)

submissions/copilot-agent-advisor/SKILL.md:37

• These Markdown tables start each row with ||, which typically renders as an extra empty column (or malformed table) in common Markdown renderers. Use a single leading | for each row/header separator to ensure the table renders correctly.

| Option | What it is | Use when |
|--------|-----------|----------|
| **Use Microsoft 365 Copilot as-is** | The built-in Copilot experience over Microsoft Graph data, no build. | Built-in capabilities + your Microsoft 365 data already cover the need; you don't need extra knowledge, custom instructions, or actions. |
| **Standard agent (declarative agent)** | Copilot configured with your instructions, knowledge, and actions, running on **Copilot's own orchestrator and foundation models**, inside Microsoft 365 apps. | You want to tailor Copilot for a focused scenario, stay inside Microsoft 365 (Teams, Outlook, SharePoint, Word), inherit Microsoft 365 compliance/RAI, and ship fast with low-code or a streamlined pro-code path. |
| **Custom agent (custom engine agent)** | A fully custom agent where you **bring your own orchestrator and models**, with custom/complex workflows, its own hosting, and multi-channel reach. | You need custom orchestration, your own/fine-tuned/domain models, deterministic complex business logic, proactive/autonomous behavior, group collaboration, delivery outside Microsoft 365, or you're integrating an existing external bot. |

submissions/copilot-agent-advisor/README.md:11

• Same table formatting issue as in SKILL.md: leading || will likely produce an unintended empty first column or break rendering. Switch to single leading | per row so the README table displays as intended.

| Option | What it is |
|--------|-----------|
| **Use Microsoft 365 Copilot as-is** | The built-in Copilot experience over your Microsoft 365 data — no build. |
| **Standard agent (declarative)** | Copilot tailored with your instructions, knowledge, and actions, running on Copilot's own orchestrator and models, inside Microsoft 365 apps. |
| **Custom agent (custom engine)** | A fully custom agent — your own orchestrator and models, custom workflows, its own hosting, and reach beyond Microsoft 365. |

submissions/copilot-agent-advisor/SKILL.md:102

• This section makes several definitive product/billing/capability claims (e.g., specific billing models, file-edit capabilities, non-transferability) while later guardrails say to verify volatile details and not fabricate. To avoid the skill outputting stale specifics when web verification isn’t available, consider rephrasing these bullets to be explicitly conditional (e.g., “Per Microsoft Learn (verify current): …”) and/or move detailed billing/capability statements under a “verify on Learn” note while keeping only the high-level decision guidance here.

- **GitHub Copilot harness — the most capable, for reasoning-heavy work.** Takes a goal, breaks it into steps, and calls tools across connectors, knowledge, MCP, and connected agents, adjusting when a step fails. Natively creates and edits Word, Excel, PowerPoint, and PDF files, supports **skills** and **memory**, and runs each task in a governed sandbox. Billed with **Copilot Credits**. Recommend it for complex, multi-step business processes that work across tools and files (e.g. an accounts-payable agent that reads invoices, matches purchase orders, and routes exceptions).
- **Standard harness — dependable and rule-based.** You define the topics, prompts, and paths so a structured, repeatable conversation or workflow responds **predictably**, drawing on your prompt library and enterprise knowledge. Billed via **standard-harness licensing**. Recommend it when the scenario is well-defined and you want consistent answers (e.g. an internal help-desk agent that answers common questions and routes simple requests).
- **Copilot chat harness — for extending Microsoft 365 Copilot Chat.** Connects your enterprise knowledge to M365 Copilot Chat so employees get grounded answers without leaving their everyday experience; publishes to internal teams. Billing is consumption-based or included in the Microsoft 365 Copilot user subscription. Recommend it when the priority is connecting people to information (e.g. an onboarding agent grounded on SharePoint knowledge).
- **Always warn:** agents aren't transferable between the **GitHub Copilot** and **standard** harnesses (different runtimes), and the harnesses bill differently, so the choice should be deliberate. You switch which harness you build on via the **New experience** toggle on the Copilot Studio home page.
- **Depth boundary.** Advise the harness at a *decision* level — which one and why. For detailed Copilot Credit / licensing estimates, capacity math, or channel-and-identity implementation checks, hand off to a dedicated harness-selection skill or to Microsoft Learn rather than computing them here. This skill focuses on the choice of whether to build and which agent type, with the harness as one downstream step.