5.1 Agent Store vs. Creating a New Agent

5.1 Agent Store vs. Creating a New Agent

Quick Answer: Copilot Agents are specialized AI assistants that extend Microsoft 365 Copilot by packaging foundation models with persistent system instructions, scoped enterprise knowledge (such as specific SharePoint sites or document libraries), functional capabilities (web search, code interpreter, image generation), and tailored conversation starters. While standard Copilot Chat is an ad-hoc conversational assistant searching broadly across a user's entire Microsoft Graph, Declarative Agents constrain and focus the AI to deliver consistent, reproducible business outputs. Organizations can choose between pre-built Agent Store agents (e.g., Researcher, Analyst, Visual Creator) for horizontal tasks, or author Custom Agents using the no-code creation wizard in Microsoft 365 Copilot or Microsoft Copilot Studio when specialized proprietary knowledge and strict compliance guardrails are required.

As organizations mature in their generative AI adoption, relying solely on open-ended, ad-hoc chat in Microsoft 365 Copilot reveals operational limitations. When individual employees craft their own prompts from scratch every day, output quality varies significantly, context drift occurs, and users must repeatedly supply background context, document references, and structural rules. To transform generative AI from an individual productivity tool into scalable, standardized enterprise business processes, Microsoft introduced Copilot Agents.

---

1. Understanding Copilot Agents: From Ad-Hoc Chat to Declarative Assistants

In standard Microsoft 365 Copilot Chat (formerly Business Chat / Microsoft 365 Chat), the system operates across an expansive, open canvas. It dynamically accesses whatever emails, chats, meetings, and documents the authenticated user has permission to see across the entire tenant. While this broad scope is powerful for general inquiries, it requires the user to manually define the Goal, Context, Source, and Expectations (the GCSE framework) in every interaction.

+-------------------------------------------------------------------------+
|                   COPILOT ARCHITECTURAL SPECTRUM                        |
+-------------------------------------------------------------------------+
|  [1. AD-HOC COPILOT CHAT]  ──► Broad Graph retrieval across all files;  |
|                                transient prompts; user defines context. |
|                                                                         |
|  [2. AGENT STORE AGENTS]   ──► Pre-built Microsoft horizontal tools     |
|                                (Researcher, Analyst, Visual Creator).   |
|                                                                         |
|  [3. DECLARATIVE AGENTS]   ──► Custom assistants built on M365 Copilot; |
|                                scoped SharePoint/OneDrive knowledge,    |
|                                persistent metaprompt, no-code authoring.|
|                                                                         |
|  [4. CUSTOM ENGINE AGENTS] ──► Pro-code, standalone autonomous bots     |
|                                built in Copilot Studio / Azure AI with  |
|                                custom LLMs & complex state machines.    |
+-------------------------------------------------------------------------+

Declarative Agents vs. Custom Engine Agents

When preparing for the AB-730 certification, it is essential to distinguish between the two primary categories of custom agents in the Microsoft ecosystem:

• Declarative Agents: Declarative agents declare customizations (instructions, knowledge sources, actions, and capabilities) that run directly on top of the native Microsoft 365 Copilot platform. They utilize Copilot's built-in foundation models (Azure OpenAI Service), inherit the enterprise security boundary, leverage the Semantic Index for Graph retrieval, and require no dedicated cloud hosting infrastructure or custom LLM orchestration logic. Business users and citizen developers can build declarative agents in minutes using Microsoft 365 Copilot Chat or Microsoft Copilot Studio.
• Custom Engine Agents: Custom engine agents are standalone, pro-code conversational bots built with custom orchestration frameworks (such as Semantic Kernel, Teams AI Library, or Azure AI Studio) and potentially alternative foundation models. They manage their own conversation state, context memory, and infrastructure, typically deployed to enterprise portals, mobile apps, or external customer-facing websites.

---

2. General Copilot Chat vs. Declarative Agents: Architectural Comparison

Understanding the operational differences between standard ad-hoc Copilot Chat and specialized Declarative Agents is fundamental to evaluating enterprise use cases.

---

3. Out-of-the-Box Agent Store Agents

Microsoft provides several pre-built, first-party agents directly within the Agent Store (accessible via the right sidebar of Microsoft 365 Copilot Chat). These agents provide immediate specialization for horizontal business tasks without requiring any authoring or configuration.

+-------------------------------------------------------------------------+
|                    FIRST-PARTY AGENT STORE AGENTS                       |
+-------------------------------------------------------------------------+
|  [RESEARCHER]      -> Multi-source web & document deep dives;           |
|                       synthesizes extensive briefs with citations.      |
|                                                                         |
|  [ANALYST]         -> Structured tabular data calculation, trend        |
|                       modeling, & Python sandbox visual charts.         |
|                                                                         |
|  [VISUAL CREATOR]  -> Visual concept ideation, diagram generation,      |
|                       & marketing graphics via Microsoft Designer.      |
|                                                                         |
|  [INTERPRETER]     -> High-fidelity multilingual translation,           |
|                       cultural tone adaptation, & localization.         |
+-------------------------------------------------------------------------+

Key Pre-Built Microsoft Agents

1. The Researcher Agent: Designed for comprehensive knowledge synthesis. When tasked with analyzing market dynamics or emerging technology standards, Researcher executes multi-step web queries, reviews attached internal whitepapers, extracts competing viewpoints, and synthesizes multi-page structured reports with footnoted citations.
2. The Analyst Agent: Built specifically for numeric and tabular data. Leveraging the integrated Code Interpreter (Python sandbox), Analyst ingests complex Excel workbooks and CSV files, cleans malformed rows, calculates statistical correlations, identifies outliers, and generates downloadable visual charts.
3. The Visual Creator / Designer Agent: Powered by Microsoft Designer and DALL-E models. It assists marketing, communications, and design teams in creating presentation imagery, conceptual illustrations, social media graphics, and iconography directly from natural language prompts.
4. The Interpreter / Language Agent: Specializes in cross-lingual communication, preserving nuanced business terminology, formatting structures, and formal corporate register across dozens of languages.

When to Adopt Agent Store Agents

Organizations should leverage Agent Store pre-built agents when:

• The business task is horizontal and universal across industries (e.g., general market research, data visualization, image generation).
• The workflow does not depend on a proprietary, restricted internal knowledge base (such as an internal employee handbook or confidential engineering wiki).
• Rapid deployment is required without authoring custom system instructions or managing ongoing agent maintenance.

---

4. Evaluating Business Requirements: When to Build Custom Agents

When standard Copilot Chat and pre-built Agent Store agents cannot meet specific organizational requirements, business units should design and deploy Custom Declarative Agents. To evaluate whether a custom agent is warranted, business analysts apply the 4-D Evaluation Framework.

+-------------------------------------------------------------------------+
|                      THE 4-D EVALUATION FRAMEWORK                       |
+-------------------------------------------------------------------------+
|  [1. DOMAIN KNOWLEDGE]       -> Does the workflow require exclusive     |
|                                 grounding in isolated internal sources? |
|                                                                         |
|  [2. DIRECTIVE CONSISTENCY]  -> Does the task mandate a rigid persona,  |
|                                 fixed schemas, or negative constraints? |
|                                                                         |
|  [3. DELIMITED CAPABILITIES] -> Must web search, code execution, or     |
|                                 image creation be strictly controlled?  |
|                                                                         |
|  [4. DISTRIBUTION & REACH]   -> Will this assistant be shared across a  |
|                                 team, business unit, or entire tenant?  |
+-------------------------------------------------------------------------+

The 4-D Evaluation Criteria Explained

1. Domain Knowledge: If an assistant must restrict its answers exclusively to a single SharePoint site (e.g., Legal Contract Templates or Clinical Trial Protocols) and avoid pulling unrelated tenant files or outdated historical drafts, a custom agent is required.
2. Directive Consistency (System Instructions): If the output must consistently adhere to a mandatory corporate tone, include legal disclaimers, format data as a specific 4-column Markdown table, or enforce strict negative boundaries (e.g., "Never provide medical or tax advice"), hardcoding these rules into a custom agent's system prompt eliminates human prompting error.
3. Delimited Capabilities & Security Controls: If corporate policy mandates that external web browsing via Bing must be disabled to prevent data leakage during sensitive M&A analysis, a custom agent allows administrators to turn off web search while keeping internal Graph grounding active.
4. Distribution & Reach: If a repeatable workflow is executed by hundreds of employees (e.g., IT onboarding or travel expense inquiries), publishing a curated custom agent with pre-defined conversation starters drives enterprise-wide adoption and productivity.

Agent Selection Decision Matrix

                                [ Business Need Identified ]
                                              │
                                              ▼
                              Is proprietary internal knowledge
                             or a fixed persona required?
                                       /             \
                                     NO               YES
                                     /                 \
                     Is it a standard horizontal        ▼
                     task (Research/Analysis)?   [ BUILD CUSTOM AGENT ]
                             /          \                  │
                           NO            YES               ▼
                          /                \     Does it require complex
                         ▼                  ▼    pro-code state logic or
                 [ Use Standard      [ Use Agent  external API writebacks?
                  Copilot Chat ]      Store Agent]     /             \
                                                     NO               YES
                                                     /                 \
                                                    ▼                   ▼
                                            [ Declarative Agent  [ Custom Engine
                                             in M365 Copilot /    Agent in Studio
                                              Copilot Studio ]    / Azure AI ]

---

5. Realistic Business Scenario: Modernizing Contoso's Procurement Workflow

The Operational Challenge

Contoso Enterprises employs 8,000 staff across 12 countries. Every month, employees submit hundreds of repetitive inquiries regarding vendor onboarding thresholds, Master Services Agreement (MSA) requirements, approved IT hardware lists, and purchase order (PO) approval matrices.

Currently, employees use standard Microsoft 365 Copilot Chat, but results are inconsistent:

• Employees often retrieve outdated 2023 procurement drafts stored in unorganized personal OneDrives.
• Inquiries regarding vendor contracts occasionally trigger ungrounded answers or hallucinated approval dollar thresholds.
• Users forget to prompt Copilot to format answers as step-by-step checklists, resulting in missing paperwork.

The Solution: Deploying a Custom Declarative Agent

The Procurement Operations Director builds a custom declarative agent called "Contoso Procurement Navigator" using the Microsoft 365 Copilot Agent Builder:

1. Knowledge Scoping: Grounding is strictly bounded to the official SharePoint Document Library: https://contoso.sharepoint.com/sites/Procurement/ApprovedPolicies2026/.
2. System Instructions: Configured with a formal persona, mandatory 4-step checklist output schemas, and strict negative constraints: "Never approve a purchase order value exceeding $50,000 without directing the user to the Senior Finance Review Board. If a vendor is not listed on the Approved Vendor Master List, state 'Vendor Unvetted' and provide the onboarding link."
3. Capabilities: Web Search is disabled (to keep all guidance strictly internal), Image Generation is disabled, and Code Interpreter is enabled (to calculate total order volume discounts).
4. Suggested Starters: Four prompt bubbles are configured ("How do I onboard a new software vendor?", "What is the PO approval threshold for hardware under $5,000?", "Check approval requirements for international contractors", "Generate vendor compliance checklist").

Outcome

Procurement support tickets drop by 62% in the first quarter, while compliance with vendor onboarding protocols reaches 98% across all regional offices.

---

Exam Warning: On the AB-730 exam, you may encounter questions asking whether declarative agents require custom Azure hosting or separate virtual machines. The answer is NO; declarative agents run entirely within the native Microsoft 365 Copilot service boundary, leveraging Azure OpenAI Service and Microsoft Graph without requiring infrastructure provisioning.

Exam Tip: Remember the distinct purpose of the two builder tabs: the Describe tab uses natural language conversational AI to generate an initial agent draft, while the Configure tab allows explicit, deterministic editing of instructions, knowledge sources, capabilities, and conversation starters.