Gokan's Studio

An Architect’s Comparison and Governance Guide

As organizations aim to become “Frontier Firms,” their strategic focus has shifted toward democratizing intelligence and obsoleting routine operational tasks. Working across the collaboration spectrum in Belgium has taught me that this transformation is driven by the shift from assistive chat to agentic workflows. In the current 2026 paradigm, an agent is defined not merely by its ability to generate text, but also by its capacity to interpret intent, access specific tools via standards such as the Model Context Protocol (MCP), and collaborate with other agents using multi-agent orchestration patterns.

Which agents are we going to discuss?

Core Agent Classification	Primary Function	Technical Host	Governance Framework
OneDrive Agents	Personal File Assistant	OneDrive	Microsoft 365 admin controls + file permissions
SharePoint Agents	Team Knowledge Base	SharePoint	SharePoint admin center + Copilot admin controls
Copilot Studio Lite	Rapid Prototyping	Microsoft 365	Microsoft 365 admin center
Copilot Studio Full	Enterprise Orchestrator	Power Platform	Full ALM + Power Platform admin + Microsoft Entra
Knowledge Agents	Content Librarian	SharePoint	PowerShell / SharePoint Admin / Purview

1 – Personal File Assistants — The OneDrive Agents

OneDrive Agents, expected to reach general availability (GA) during 2026, are the most localized layer of the agentic collaboration era. These are “Personal Assistants” designed to eliminate friction in retrieving information from individual document silos. Architecturally, a OneDrive Agent is a declarative instance of Microsoft 365 Copilot, scoped to a specific set of files or folders selected by the user. The core value proposition of a OneDrive Agent is its high precision and low hallucination rate, achieved through narrow grounding. When creating the agent, the wizard shows a warning about scope limitations:

When a user selects ten technical specifications for a project and instantiates an agent, the system creates a temporary index of those specific documents. Unlike the broader Copilot experience, which may pull in “noise” or “hallucinations” from the entire Microsoft Graph, the OneDrive Agent is restricted to the chosen context. This makes it an ideal tool for:

• Project contextualization
• Research synthesis across dozens of PDFs
• Meeting preparation using a folder of historical minutes

Despite their utility for individual productivity, OneDrive Agents are not designed for organizational scale. They have very low extensibility—they cannot connect to external APIs or call custom connectors; they are limited to reading selected content and using built-in Copilot capabilities. Furthermore, they are essentially siloed: an agent created by User A to analyse a private project folder cannot be easily shared or repurposed as an enterprise-wide tool without migrating its logic to a more robust platform such as Copilot Studio.

From an administrative perspective, governance is managed through the Microsoft 365 admin center. Admins can see which users are creating agents, but the agent’s actual “Knowledge Set” is governed by existing OneDrive file permissions. If a user loses access to a file, the agent’s ability to reference that file is immediately revoked, ensuring the principle of least privilege is maintained without manual IT intervention.

1.1 – Use Cases

• Project Contextualization: “Onboard me to Project Alpha using these 10 specs and meeting notes.”
• Research Synthesis: Summarizing themes across 20+ PDF research papers without hallucinations from outside sources.
• Meeting Prep: Querying a specific folder of past minutes to prepare for an upcoming review.

1.2 – Key Attributes

Attribute	Details
Licensing	Included with Microsoft 365 Copilot ($30/user/month on annual terms)
Scalability	Low — scoped to individual user files or shared folders. Not designed for organization-wide deployment.
Extensibility	Very Low — restricted to selected files; no custom APIs or external data connectors.
Pros	Zero friction creation; high accuracy from narrow scope; inherits existing file permissions instantly
Cons	Manual updates if sources change; siloed (cannot integrate with other systems)
Admin Capabilities	Minimal — primarily user-driven; admins manage via standard OneDrive/SharePoint file policies

2 – Collective Intelligence — SharePoint Agents

SharePoint Agents serve as the conversational layer of the corporate intranet. In 2026, every SharePoint site is provisioned with a default agent, but site owners—typically department leads or project managers—can create custom “Declarative Agents” tailored to specific libraries. One day Lists will be there too 😊

These agents are designed to move beyond the never-working-always-needs-a-developer Search paradigm, providing a non-actionable “Chat” experience that understands the nuances of departmental knowledge. Deploying a SharePoint Agent is a no-code process that site owners can perform without IT assistance.

The technical maturity of SharePoint Agents in 2026 is characterized by metadata-aware retrieval. Unlike traditional RAG systems that only look at text chunks, SharePoint Agents can interpret site metadata. For instance, if an agent is asked for “All active contracts expiring this quarter,” it uses the structured metadata columns (e.g., “Status” and “Expiry Date”) to filter the document library before performing text analysis. This capability is significantly enhanced when a content enrichment process has previously been deployed to structure and clean that data.

2.1 – Governance and Content Safety

A critical governance capability in 2026 is the ability to exclude specific sites and files from AI grounding, even while they remain visible to users in search. When an administrator applies these content safety controls on a site, it adds metadata flags that prevent Microsoft 365 Copilot and all declarative agents from using those files for grounding, even if the user has read access. This is essential for protecting confidential executive communications or sensitive intellectual property that should remain searchable by humans but invisible to large language models. As you can see, there is a little “PRO” sign next to each option, which means that you need to have a SharePoint Advanced Management license (which is free when you have at least one Copilot license).

2.2 – Use Cases

• Departmental Help: “What is the travel policy for international flights?” (scoped to HR Policy library)
• Onboarding Buddy: An agent that answers questions solely from a “New Hire” site.
• Document Retrieval: Finding specific contracts or invoices based on metadata queries.

2.3 – Key Attributes

Attribute	Details
Licensing	Can interact via M365 Copilot OR pay-as-you-go consumption billing
Scalability	Medium — great for team or department scale; shareable via Teams.
Extensibility	Medium — can be extended via Copilot Studio to add Power Automate flows for simple actions.
Pros	Context-aware; automatically grounded in site’s trusted content; ease of sharing via Teams or Outlook; no-code creation by site owners
Cons	Scope primarily limited to SharePoint content unless upgraded to Copilot Studio Full
Admin Capabilities	Strong — site-level opt-in/opt-out; respects all SharePoint permissions and Microsoft Purview labels

3 – The Role of Copilot Studio Lite (Agent Builder)

Copilot Studio Lite, often referred to as “Agent Builder” in the Microsoft 365 Copilot interface, democratizes agent creation. It enables non-technical employees to build functional agents using natural-language prompts rather than traditional configuration menus. This tool is designed for rapid prototyping—creating specialized personal productivity tools, such as a “Writing Coach” or a “Simple Q&A Bot,” for a specific team event.

3.1 – How It Works

The “Lite” experience is essentially a managed conversational authoring interface. A user describes the agent’s purpose (e.g., “Create an agent that helps me critique my emails to be more professional”), uploads relevant training documents, and the system automatically generates the underlying instructions and grounding configurations. These agents automatically benefit from the latest LLM capabilities (currently leveraging GPT-4/GPT-5-series models), ensuring high-quality responses with zero technical overhead.

3.2 – Technical Guardrails

To maintain enterprise safety, Studio Lite agents are sandboxed. They are limited to standard Microsoft 365 knowledge sources (Web, SharePoint, Microsoft Graph) and cannot call custom APIs or write back to external systems. Importantly for IT budgets, these agents do not consume the tenant’s Dataverse storage entitlement, as they are processed by the Copilot Studio service as lightweight templates rather than full-scale application resources.

The most significant strategic value of Studio Lite is the “One-Click Upgrade“. If a prototype built in the Lite interface proves highly successful, a developer can use the “Copy to Copilot Studio” action to migrate the agent to the “Full” experience. This preserves the initial instructions and knowledge grounding while unlocking professional-grade features, including multi-step logic, variables, and 1,400+ system connectors.

3.3 – Use Cases

• Personal Productivity: A “Writing Coach” agent instructed to critique emails in a specific tone.
• Simple Q&A: Quick bot for a specific event or short-term initiative.
• Prototyping: Testing an agent concept before asking IT to build a full enterprise version.

3.4 – Key Attributes

Attribute	Details
Licensing	Included with Microsoft 365 Copilot ($30/user/month)
Scalability	Low/Medium — best for personal use or small group sharing.
Extensibility	Low — limited to standard knowledge sources (Web, SharePoint, Graph); no custom API access.
Pros	Fastest time-to-value: build an agent in < 1 minute using natural language; empowers non-technical staff to solve their own problems
Cons	Limited transparency: you cannot easily see or edit the complex backend orchestration; no formal ALM/versioning lifecycle
Admin Capabilities	Basic — tenant-level on/off toggles in Microsoft 365 admin center

4 – Autonomous Enterprise Orchestration — Copilot Studio

Copilot Studio is the professional developer’s platform for building autonomous agents. It is a comprehensive low-code/pro-code environment that supports complex logic, conditional branching, and deep integration with external systems like SAP, Salesforce, and Oracle. While the other agent types are primarily assistive (reactive to user queries), Studio Full agents are increasingly agentic (proactive in executing workflows and making decisions).

4.1 – Multi-Agent Orchestration and Coordination

In the 2026 architecture, Studio agents function as orchestrators. Using multi-agent orchestration patterns, a primary agent can delegate sub-tasks to specialized remote agents. For example, an IT Service Desk agent might receive a request to “Onboard a new employee” and then:

1. Delegate account creation to an “Identity Agent“
2. Delegate hardware procurement to a “Logistics Agent“
3. Delegate training enrollment to an “HR Agent“
4. Consolidate status updates and notify the manager

This horizontal scaling allows organizations to build an “Agentic Mesh” where specialized bots collaborate to achieve complex goals, rather than building a single, monolithic bot that tries to handle every process.

4.2 – Operational Governance and ALM

For the IT professional, Copilot Studio is the only agent environment in this lineup that supports a traditional software development lifecycle (SDLC). This includes:

• Application Lifecycle Management (ALM) with Dev/Test/Prod pipelines
• Version control with concurrent editing and rollback
• Deep telemetry to monitor agent performance and quality over time

This level of control is necessary for customer-facing bots or high-value internal automations where the risk of error carries significant business impact.

4.3 – Use Cases

• IT Service Desk: An agent that can troubleshoot, open a ticket in ServiceNow, and reset a password.
• HR Automation: End-to-end leave application handling (checking balance in ERP, updating calendar, notifying manager).
• Customer Facing: External bots on websites for customer support.

4.4 – Key Attributes

Attribute	Details
Licensing	Separate cost — requires Copilot Studio license (tenant license + user licenses). Alternatively, Azure pay-as-you-go
Scalability	High — enterprise-wide; supports millions of conversations and complex branching logic.
Extensibility	Exceptional — 1,400+ connectors (via Power Platform), custom APIs, Azure Logic Apps, Code execution (Python, etc.)
Pros	Full control over every step of conversation and logic; can read AND write to almost any external system; fully branded experience (web, mobile app, custom channels)
Cons	Higher cost and steeper learning curve; requires developer or power-maker skills
Admin Capabilities	Very Strong — Application Lifecycle Management (ALM) with Dev/Test/Prod pipelines; deep analytics via Power Platform admin and Purview; comprehensive DLP policies

5 – The SharePoint Layer — Content Enrichment and Knowledge Organization

In late 2025 and early 2026, Microsoft is expanding content preparation and enrichment capabilities designed to ensure that organizational documents are AI-ready. Literally translating to bye-bye Syntex, welcome Knowledge Agent. These tools work in the background to prepare the organization’s content for consumption by all other agents.

5.1 – The “GIGO” Problem and Content Preparation

The core challenge is the “Garbage In, Garbage Out” (GIGO) problem that plagues many AI rollouts: if your data is messy, unstructured, or missing context, all of your agents will inherit that problem.

Content enrichment capabilities (increasingly available in Document Processing / SharePoint Premium / Syntex and M365 Copilot) can automatically tag documents with organizational taxonomy, extracting metadata like “Invoice Number,” “Total Amount,” “Vendor Name” from thousands of unsorted files

Beyond metadata, these capabilities can also act as an automated content management tool. They can:

1. Identify content gaps — Analyze search behavior and usage patterns to see what information users are looking for but failing to find.
2. Fix broken links — Proactively identify dead links within a site.
3. Manage lifecycle — Flag stale or duplicate content for review or archiving, ensuring the AI is only grounded in authoritative, up-to-date information.

5.2 – Use Cases

• Metadata enrichment: Automatically tagging thousands of uploaded contracts with “Expiry Date,” “Vendor Name,” and renewal status.
• Governance: Identifying and flagging stale or duplicate content that should be excluded from search and AI grounding.
• Reliability boosting: Improving the “grounding quality” of Copilot by ensuring the underlying data structure is clean and well-organized.

5.3 – Key Attributes

Attribute	Details
Licensing	Typically tied to Document Processing (Syntex) pay-as-you-go services or emerging M365 Copilot add-ons, models are still evolving in early 2026, but LLM is the preferred way.
Scalability	Organization-wide — runs against entire SharePoint sites, libraries, and tenant-level document collections.
Extensibility	Medium — focuses on SharePoint taxonomy, metadata columns, and content tagging; integration with external systems is limited.
Pros	Significantly improves the success rate of all other agents by fixing data quality; removes the manual drudgery of manual tagging and organization; increasingly automated via AI
Cons	Not a chatbot — it is a tool for IT managers and site owners to prepare content, not a user-facing agent
Admin Capabilities	Strong — governance-focused; controls what content is “authoritative” for AI grounding; integrates with Purview and DLP

Identity, Security, and Governance at Scale

AI agents aren’t “just chatbots”. In our ecosystem, an agent can access enterprise knowledge data, connect to other business systems, applications, and take actions (e.g., create tickets, send emails, call APIs, update records). That enables agents to operate autonomously, and they deserve Zero Trust controls across identity, data, runtime, and network.

AI agents built on Microsoft platforms (Copilot Studio, Microsoft 365 Agents/Agent Builder) introduce new productivity opportunities but also new risks. Microsoft’s own agent guidance frames security as a lifecycle discipline spanning inception, deployment, and ongoing operations.

This part of the article will provide a client-ready security strategy for Microsoft AI agents based on five pillars:

1. Identity & Access — govern agent identities with Microsoft Entra Agent ID / Agent 365
2. Platform Governance — enforce Copilot Studio and Power Platform policies (data policies/DLP enforcement)
3. Data Security & Compliance — apply Microsoft Purview controls and DSPM for AI
4. Runtime Threat Protection — block malicious actions at tool invocation via Microsoft Defender
5. Network Security — extend web controls to agent traffic using Secure Web & AI Gateway (Global Secure Access)

Layer 1 — Identity & access: treat each agent as a governed identity

The first step is to start by modeling agents as workloads that need their own identity. Microsoft Entra Agent ID introduces agent identities (a specialized identity construct in Entra) and a platform for lifecycle + access management at scale. The key idea is to have an agent identity that is described as a special service principal in Entra ID with consent-based token acquisition

Our Recommendation:

• Give each production agent a dedicated agent identity (and ownership by a group, not an individual).
• Apply Conditional Access patterns for agent access paths (where applicable) and enforce strong sign-in controls for makers/admins managing agents.
• Run least-privilege reviews for every connector and action an agent can execute.

If you’re adopting Agent 365, Microsoft’s direction is: each agent gets an Entra Agent ID and becomes manageable/observable via the Microsoft 365 admin experience.

An agent is only as secure as the identity behind it.

Below is a practical way to understand an agent’s architecture, how it maps to Microsoft Entra ID, and what you must secure if you want agents to be safe in production.

Think of an agent as a system with four key building blocks:

• Model: The AI model, General, Custom, Multimodal
• Orchestration layer: The loop that processes input, reasons, and decides the next action.
• Memory: Memory gives agents up-to-date context so responses stay accurate and relevant.
• Tools: Tools let agents interact with systems and extend capabilities (web search, databases, APIs, files, integrations).

In short: Model + Orchestration + Memory + Tools = an agent that can act.

Why does the agent needs an identity in Entra ID

The moment your agent needs to access enterprise resources—like a web service, an internal API, Microsoft Graph call, or a third-party SaaS—it requires authentication. That’s where Microsoft Entra ID comes in.

Your agent will typically be represented by an agent identity in Entra ID (commonly implemented as an app/workload identity). This identity can:

• Request access tokens from Entra ID
• Use those tokens to call protected web services / APIs

So, yes: your agent has an identity, and that identity becomes a high-value target.

Layer 2 — Platform governance: lock down Copilot Studio with data policies

Copilot Studio provides governance and DLP-style controls via the Power Platform admin center, including the ability to restrict publishing, generative features, and data interactions. The most important control here is Data policies: they govern how agents connect to data/services inside and outside your org. Microsoft also notes that (since early 2025) data policy enforcement is in effect for all tenants — so governance is not optional; it’s foundational.

Our recommendation:

• Separate Dev / Test / Prod environments.
• Start with deny-by-default in Prod for:
  • HTTP / generic web calls
  • unknown connectors
  • external knowledge sources
• Allow exceptions through a controlled process (ticket + approval + periodic recertification).

Layer 3 — Data protection & compliance: Purview guardrails for prompts, answers, and knowledge

Agents amplify data risk in new contexts. Microsoft Purview provides specific guidance on managing data security and compliance for interactions with agents built into Copilot Studio. For broader posture management, Microsoft also highlights DSPM for AI (preview) as a central hub for securing data for AI apps and proactively monitoring AI use.

Our Recommendation:

• Implement Sensitivity Labeling and align agent knowledge sources with label-based access decisions.
• Deploy DLP policies (including Copilot/AI interaction coverage where relevant) to reduce accidental exposure.
• Turn on auditing and build an investigation path (eDiscovery/retention if required).

Layer 4 — Runtime protection: block attacks at the moment the agent acts

The highest-risk moment is tool invocation (when the agent executes an action or calls an external system). Microsoft Defender now provides real-time protection during agent runtime for Copilot Studio agents, with configuration steps in Defender for Cloud Apps settings.

Our Recommendation:

• Enable Defender runtime protection for Copilot Studio agents (and integrate alerts into your SOC flow).
• Create incident playbooks for:
  • prompt injection attempts
  • suspicious tool calls
  • abnormal access patterns
• Review blocked actions weekly and tune allow lists/guardrails.

Layer 5 — Network controls: Secure Web & AI Gateway (Global Secure Access)

Identity controls don’t cover everything if your agent can browse anywhere or download/relay files. Microsoft documents Secure Web & AI Gateway for Copilot Studio agents via Global Secure Access, enabling enforcement through a globally distributed proxy and policies such as web content filtering and threat intelligence filtering.

Our Recommendation:

• For any agent with outbound web/tool access, enforce:
  • web content filtering
  • threat intel filtering
  • (where relevant) network file filtering
• Run with strict categories in Prod; open up only when business justified.

Bonus – Secure Your Microsoft AI Agents: A 10-Point Governance Checklist

This is a practical set of controls I’ve used to reduce risk—like data leakage, over-permissions, and shadow agents—while helping teams scale AI.

Conclusion: Building the foundations of the Frontier Firm

I can’t say whether the Frontier paradigm will be a success, as many of Microsoft’s initiatives have failed epically (Microsoft relies primarily these days on its flagship tool and programs), but the 2026 Microsoft AI agent ecosystem provides the most comprehensive framework for organizational automation to date. By differentiating between file-centric OneDrive Agents, site-centric SharePoint Agents, rapid-prototyping Copilot Studio Lite, and process-orchestrating Copilot Studio Full agents, IT professionals can align the right tool with the right business need.

So, what should you use and when?

Scenario	Recommended Tool	Why
“I just need to chat with these 5 PDFs for my project.”	OneDrive Agents	Zero setup; keeps project context isolated; immediate access control inheritance
“My team needs a bot to answer questions about our HR policies on our intranet.”	SharePoint Agent	Lives where the team works; site owners can maintain it without IT; easy sharing via Teams
“I want to create a quick ‘Email Helper’ bot for myself or my immediate peers.”	Copilot Studio Lite	Fast; free (with Copilot license); describe in natural language; one-click upgrade path if needed
“We need a bot that can look up an order in SAP and process a refund.”	Copilot Studio	Requires external API connection and complex “write” actions that Lite cannot handle; multi-agent orchestration available
“Our Copilot answers are wrong because our SharePoint data is a mess.”	SharePoint KA	Use this to clean and tag data first, so the other agents work correctly; improves grounding across all agents
“We’re building a mission-critical customer support bot with SLA requirements.”	Copilot Studio	Full ALM, monitoring, scaling, and multi-region deployment; enterprise integrations; DLP and governance

In my humble experience and based on what we’ve discovered, the successful “Frontier Firm” will not deploy these agents in isolation. Instead, they will:

1. Ensure data quality by deploying content enrichment on high-value libraries
2. Leverage multi-agent orchestration patterns to build an Business Process where agents collaborate
3. Govern the entire estate through Microsoft Entra Agent IDs, Purview DLP, and admin center observability
4. Rationalize and Optimize costs using a hybrid seat-and-credit licensing strategy

In this agentic era, the competitive advantage belongs to those who can orchestrate intelligence at scale.

---

Feedback & Contributions

This is a community-driven initiative, and I may be mistaken. If something is not accurate, or if you want to add key information with proper attribution, please share feedback at [gokan@neoxy.be] or [seyfallah.t@gmail.com] for Purview.

Hope this helps!

Renewed Revolution!