MCP, Computer Use & Agent Behaviours

New frontiers: MCP, Computer Use, and smart behaviours

Model Context Protocol (MCP) in Copilot Studio

MCP is an open protocol that standardises how AI agents connect to external tools and data. Instead of building a custom connector for every integration, you connect to an MCP server that exposes tools, resources, and prompts in a standard format.

How MCP works in Copilot Studio:

1. MCP server — an external service that exposes tools (functions the agent can call), resources (data the agent can read), and prompts (pre-built prompt templates)
2. Connection — Copilot Studio connects to the MCP server endpoint
3. Discovery — the agent discovers available tools and resources from the MCP server
4. Invocation — during a conversation, the agent calls MCP tools when relevant to the user’s request

MCP vs custom connectors vs Copilot connectors

Feature	Architecture	Best For	Effort
MCP connections	Open protocol — connect to any MCP-compatible server. Agent discovers tools dynamically	When MCP servers already exist for the tools you need, when connecting to open-source tool ecosystems, or when you want a single integration pattern across many tools	Low if MCP server exists. Medium if you need to build one
Custom connectors	Point-to-point REST API integration — you define the schema, authentication, and endpoints	When the external system has a REST API but no MCP server, when you need fine-grained control over the API interaction	Medium — define connector schema, configure auth, map data
Copilot connectors	Prebuilt Microsoft connectors in the connector gallery — ready to use with authentication configured	When a prebuilt connector already exists for the target system (Salesforce, ServiceNow, SAP, etc.)	Low — select from gallery, configure credentials

💡 Scenario: Natalie implements MCP for a client's inventory system

Natalie’s client has a custom inventory management system built in-house. Ravi (senior developer) has already built an MCP server that exposes three tools:

• check_stock(product_id) — returns current inventory levels
• find_alternatives(product_id) — finds substitute products when stock is low
• create_reorder(product_id, quantity) — creates a purchase order

Copilot Studio integration:

1. Natalie adds the MCP connection in Copilot Studio, pointing to the MCP server endpoint
2. The agent automatically discovers the three tools
3. When a sales rep asks “Do we have 500 units of part X in stock?”, the agent calls check_stock and responds with current levels
4. If stock is low, the agent proactively calls find_alternatives and offers substitutes
5. The rep can say “Order 200 more of part X” and the agent calls create_reorder

Why MCP over a custom connector? The MCP server already exists, the tool discovery is automatic, and if Ravi adds new tools later (like track_shipment), the agent discovers them without reconfiguration.

💡 Important: MCP requires generative orchestration

MCP connections in Copilot Studio require generative orchestration to be enabled on the agent. Classic topic-based agents cannot call MCP servers directly — MCP tools are invoked through the generative orchestrator, which decides when and how to use them based on the conversation context.

When to prefer MCP: Choose MCP when you want centrally managed, reusable tools that multiple agents can share. An MCP server built once can serve many agents — unlike custom connectors that are configured per agent.

Computer Use in Copilot Studio (preview)

Computer Use gives agents the ability to interact with applications and websites by seeing the screen and performing actions — clicking buttons, typing into fields, navigating menus, and reading displayed content.

Preview notice: Computer Use is currently in preview and available in the US region only. It requires generative orchestration to be enabled. There are two hosting options: hosted machines (Microsoft-managed VMs — simpler setup, Microsoft handles infrastructure) and bring-your-own machines (self-hosted VMs — more control, you manage the infrastructure). Consider latency implications for production use, as each screen observation and action cycle adds delay.

Use cases:

Scenario	Why Computer Use	Alternative
Automating data entry in a legacy ERP with no API	No programmatic access — the only interface is the UI	Build an API layer (expensive, time-consuming)
Testing web applications	Navigate pages, fill forms, verify content	Dedicated testing tools (Playwright, Selenium)
Extracting data from desktop applications	Information is on-screen but not accessible via APIs	Manual extraction or custom screen scraping
Training and onboarding demonstrations	Show step-by-step how to use an application	Recorded videos or documentation

Design considerations for Computer Use:

• Latency: Screen observation and action take time — expect 2-5 seconds per step. Not suitable for real-time interactions
• Accuracy: Complex UIs with dynamic elements can confuse the agent. Simple, consistent layouts work best
• Security: The agent sees everything on screen — ensure sensitive data is masked or the session is scoped appropriately
• Reliability: UI changes (button moves, new pop-up) can break automation. Build error handling and retry logic
• Scope: Define exactly which applications and actions the agent is allowed to interact with

💡 Scenario: Ravi builds a Computer Use agent for legacy ERP data entry

Ravi Krishnan (Cloudbridge Partners) automates data entry into a client’s legacy ERP system that has no API. The client’s warehouse team currently copies shipping data from emails into the ERP manually — 80 entries per day.

Computer Use agent design:

1. Agent receives shipping notification data (structured from an email processing step)
2. Opens the legacy ERP application
3. Navigates to the “New Shipment” form
4. Fills in fields: tracking number, carrier, ship date, destination, weight
5. Clicks “Save”
6. Verifies the confirmation message appears
7. Logs the result for audit

Guardrails:

• Agent can only interact with the ERP’s shipment entry form — no access to other modules
• Every entry is logged with before/after screenshots for audit
• If the agent encounters an unexpected screen (error dialog, login prompt), it stops and alerts the warehouse manager
• A human reviews a random 10% sample daily

Natalie presents the ROI: 80 entries at 3 minutes each = 4 hours of manual work per day. The agent handles 90% automatically, saving 3.6 hours daily.

Deep reasoning (preview)

Deep reasoning enables enhanced multi-step analysis for complex tasks. When enabled, the agent methodically breaks problems into steps and plans its approach before responding. This produces higher-quality answers for complex queries but is slower and costlier than standard orchestration.

When to enable deep reasoning:

• Multi-step calculations or analyses
• Complex decision-making with multiple variables
• Tasks that require comparing options and making trade-offs
• Situations requiring careful, methodical problem decomposition

When NOT to enable deep reasoning:

• Simple Q&A from a knowledge base
• Straightforward topic routing
• Tasks where speed matters more than depth

Note: Deep reasoning improves the agent’s internal analysis quality. It does not expose its reasoning chain to the end user — the user receives a final answer, not the intermediate steps.

Voice mode

Voice mode enables spoken interaction — the agent listens to speech, processes it, and responds with synthesised voice. Designed for phone and IVR (Interactive Voice Response) scenarios.

Design considerations for voice agents:

• Latency is critical — voice users expect fast responses. Keep prompts concise and responses under 30 seconds
• No visual fallback — you can’t show tables, links, or images. Design responses for audio only
• Confirmation patterns — “I heard you say you want to cancel order 1234. Is that correct?”
• Interruption handling — users may interrupt mid-response. Design for barge-in support
• Escalation — always provide a “speak to a human” option

Flashcards

Question

What is MCP (Model Context Protocol) and how does it work in Copilot Studio?

Click or press Enter to reveal answer

Answer

MCP is an open protocol that provides a standard interface for agents to connect to external tools, resources, and prompts. In Copilot Studio, you add an MCP connection to a server endpoint — the agent automatically discovers available tools and can invoke them during conversations. It replaces the need for individual custom connectors when MCP servers exist.

Click to flip back

Question

What is Computer Use in Copilot Studio?

Click or press Enter to reveal answer

Answer

Computer Use (preview, US region only) enables agents to interact with applications and websites visually — observing the screen and performing actions like clicking, typing, and navigating. It bridges the gap for legacy systems with no API. Requires generative orchestration. Hosting options: hosted machines (Microsoft-managed) or bring-your-own machines (self-hosted). Design considerations: latency (2-5 seconds per step), accuracy on complex UIs, security (agent sees the screen), and reliability (UI changes can break automation).

Click to flip back

Question

When should you enable deep reasoning for a Copilot Studio agent?

Click or press Enter to reveal answer

Answer

Enable deep reasoning (preview) for complex tasks: multi-step calculations, decision-making with multiple variables, and comparing options with trade-offs. Deep reasoning improves internal analysis quality but is slower and costlier than standard orchestration — the user receives a final answer, not intermediate steps. Don't enable it for simple Q&A or straightforward routing.

Click to flip back

Question

What are the key design considerations for voice-enabled agents?

Click or press Enter to reveal answer

Answer

Latency (voice users expect fast responses), no visual fallback (design for audio only — no tables or links), confirmation patterns (repeat back what the user said), interruption handling (barge-in support), and always provide escalation to a human agent.

Click to flip back

Question

When should you choose MCP over a custom connector in Copilot Studio?

Click or press Enter to reveal answer

Answer

Choose MCP when: an MCP server already exists for the target tool, you want to connect to an open-source tool ecosystem, or you want dynamic tool discovery (agent finds new tools automatically). Choose custom connectors when: no MCP server exists, you need fine-grained control over API interactions, or the external system only has a REST API.

Click to flip back

Knowledge check

Knowledge Check

A client has a 15-year-old desktop HR application with no API. HR staff spend 2 hours daily copying new hire data from a spreadsheet into the application's forms. Natalie proposes using Computer Use to automate this. What is the MOST critical design element she must address?

ALatency — Computer Use is too slow for data entryBVoice mode — HR staff should be able to speak the data instead of using a spreadsheetCError handling — define what the agent does when it encounters an unexpected screen (error dialog, login prompt, UI change) and ensure human oversight of resultsDReasoning mode — the agent needs to understand the HR data before entering it

Check Answer

Knowledge Check

Ravi has built an MCP server that exposes 5 tools for a client's custom CRM. He adds the MCP connection in Copilot Studio. Two weeks later, he adds 2 new tools to the MCP server. What does the Copilot Studio agent need?

AA new custom connector for each of the 2 new toolsBManual reconfiguration of the MCP connection to register the new toolsCNothing — MCP supports dynamic tool discovery, so the agent automatically finds the new tools on the serverDA complete rebuild of the agent to accommodate the new capabilities

Check Answer

Knowledge Check

Jordan is designing a patient appointment scheduling agent for CareFirst Health. Patients will call a phone number, speak their scheduling request, and the agent should book, reschedule, or cancel appointments. Which combination of agent behaviours should Jordan enable?

AComputer Use and voice mode — to interact with the scheduling system by screenBDeep reasoning only — the scheduling logic is the hardest partCVoice mode only — the agent just needs to understand speechDVoice mode and deep reasoning — for complex scheduling logic and spoken interaction

Check Answer

🎬 Video coming soon

Next up: M365 Agents: Teams, SharePoint & Sales/Service in M365 Copilot — optimising agent design across Microsoft 365, configuring Sales and Service in M365 Copilot, and leveraging the AI hub in Power Platform.