DE EN
Discuss your AI use case

← Back to overview

Knowledge

Copilot Studio in detail

For everyone who wants the details: how an agent is built, where it sits in your IT landscape, which models are available and how control works. Every statement with source and date.

Anatomy of an agent

What is an AI agent made of?

An AI agent has four building blocks: instructions, knowledge, tools and a language model. Governance provides the frame. Copilot Studio brings all five together on one platform – turning an idea into an agent your IT team can stand behind.

How a task flows through the agent

  1. TriggerQuestion, email, call, new record, schedule
  2. UnderstandCapture goal and context
  3. PlanChoose steps, knowledge and tools
  4. ActExecute in your systems
  5. ControlApproval, escalation, audit log
01

Instructions

Instructions and reusable skills define what the agent should do – and what it must not do.

  • Role, goal and boundaries
  • Tone and approval rules
  • Your proposal policy as a skill
02

Knowledge

The agent works from approved sources and cites them in its answers.

  • SharePoint, Dataverse, files
  • Work IQ (preview): email, calendar, Teams
  • Foundry IQ (preview), website
03

Tools

Tools turn knowledge into work: the agent reads from and writes to your systems.

  • Connectors to SAP, Salesforce, ServiceNow, Dynamics 365
  • Workflows and MCP servers
  • Computer use, other agents via A2A
04

Model

The language model brings understanding and judgment – your choice for each agent.

  • OpenAI (GPT)
  • Anthropic (Claude)
  • Mistral and models from Microsoft Foundry

Frame: governance

Every agent has its own identity and clear permissions. Cost, quality and impact are measured.

  • Microsoft Entra Agent ID
  • Least-privilege permissions
  • Data policies (DLP)
  • Evaluations and monitoring
  • Microsoft Agent 365
  • Credit budgets

And what is a harness?

The harness is the runtime between agent and model. It decides when to call the model, which context goes along and which tool gets used.1

Think of the model as a new colleague with knowledge and judgment. The harness is everything she needs to get work done: access, tools, rules and a team.

  • GitHub Copilot harnessFor multi-step processes: breaks a goal into steps, works with Word, Excel, PowerPoint and PDF, uses skills and memory (preview). Not to be confused with the GitHub Copilot service – your data does not go there.
  • Standard harnessFor rule-based agents and well-structured flows with predictable behavior: topics and agent flows.
  • Copilot chat harnessExtends Microsoft 365 Copilot Chat with enterprise knowledge – internal use only.
Architecture

Where does Copilot Studio sit in your IT landscape?

Copilot Studio is the orchestration layer between the channels people work in and the systems where the work actually happens. Knowledge, tools and governance are built once – and shared by every agent.

Channels

  • Teams
  • Microsoft 365 Copilot
  • SharePoint
  • Website
  • WhatsApp
  • Mobile app
  • Phone

Copilot Studio

  • Agents and connected agents
  • Harness
  • Models
  • Knowledge
  • Workflows

Tools and open standards

  • Connectors
  • MCP
  • A2A
  • REST API
  • Computer use

Your systems

  • ERP: SAP, Abacus, Dynamics 365
  • CRM
  • Ticketing
  • Documents
  • Industry software
  • Legacy systems without an API

Governance

  • Entra Agent ID
  • DLP
  • Evaluations
  • Monitoring
  • Agent 365
  • Credit budgets
Diagram: channels, Copilot Studio, tools and systems within a governance frame. One agent, many channels: the same agent can work in Teams, on your website and on the phone.
One studio, many models

Which AI models can I use in Copilot Studio?

In Copilot Studio, you choose the language model for each agent – from OpenAI, Anthropic or Mistral – and bring in models from Microsoft Foundry for individual steps. The model fits the task, not the other way around.

General

Fast and cost-efficient: drafts, summaries, FAQ answers, simple actions.

Deep

Deliberate, multi-step reasoning: contracts, policies, long documents, analysis across systems.

Auto

Chooses dynamically per request: mixed workloads, for example in a help desk. Currently preview only, on the standard harness.

Model picker: which model fits your task?

What should the agent take on?

Standard answers, drafts, summaries

RecommendationGeneral category

Examplese.g. GPT-5.5 Chat

Data noteFast and cost-efficient. In Europe, GPT-5.5 Chat is processed in-region.

Contracts, policies, long documents, multi-step analysis

RecommendationDeep category

Examplese.g. Claude Opus 5 (GitHub Copilot harness)

Data noteStrongest reasoning. In Europe, Claude models are cross-geo and may process data outside the region – use them only for data that allows it.

Mixed requests in a help desk

RecommendationAuto category (preview)

ExamplesGPT-5 Auto (preview, standard harness only) – routes each request between fast and deliberate models

Data noteGood for unpredictable requests. As a preview, however, it isn't meant for production – until it's generally available, we deliberately choose a General or Deep model.

Sensitive data, in-region processing required

RecommendationA model without cross-geo in Europe

ExamplesAs of September 2026: GPT-5 Chat, GPT-5.5 Chat and GPT-6 Astra

Data noteProcessing stays in the region. The admin setting “Move data across regions” stays off.

A specialized task in a single prompt step

RecommendationYour own model from Microsoft Foundry

Examplesmore than 1,800 models in the catalog, e.g. Llama or DeepSeek

Data noteChat completion models only; the GPT-5 family is currently not available for bring your own model.

Models in Copilot Studio (as of September 2026)

Models on the GitHub Copilot harness, Europe region (except UK) with category, status and data note
ProviderModelCategoryStatusData note
OpenAIGPT-5 ChatGeneralGAin-region
OpenAIGPT-5.5 ChatGeneralGAin-region
OpenAIGPT-6 AstraGeneralGAin-region
AnthropicClaude Sonnet 4.6GeneralGAcross-geo
AnthropicClaude Sonnet 5GeneralGAcross-geo
AnthropicClaude Fable 5GeneralGAcross-geo
AnthropicClaude Fable 5.1GeneralGAcross-geo
AnthropicClaude Opus 4.8DeepGAcross-geo
AnthropicClaude Opus 5DeepGAcross-geo
MistralMistral Medium 3.5GeneralExperimentalin-region

Models change constantly. We keep this table current and plan model changes as part of governance. On the standard harness, GPT-5.5 Chat is the default model; the selection differs in part.2

Standard harness: default GPT-5.5 Chat; GA: GPT-4.1, GPT-5 Chat, Claude Sonnet 4.6, Claude Opus 4.6 and 4.7 (Claude is cross-geo in Europe); preview: GPT-5 Reasoning (Deep) and GPT-5 Auto (Auto); experimental: Mistral Medium 3.5 and, in US early access environments only, GPT-5.3 Chat, GPT-5.4 Reasoning, GPT-5.5 Reasoning and Grok 4.1 Fast. GPT-4o and Claude Sonnet 4.5 are retired.3

External models (Anthropic, Mistral) must be enabled by an admin in both the Power Platform admin center and the Microsoft 365 admin center. Anthropic operates as a Microsoft subprocessor.

Not everything has to come from Microsoft

Copilot Studio is built to be open: agents speak open standards, reach third-party systems and work with models and agents from other vendors.

  • open standard

    MCP – Model Context Protocol

    Systems describe their own tools and data. Connect once and every agent can use them – prebuilt MCP servers exist for Dataverse, Outlook, Salesforce and Jira, among others.4

  • open standard

    A2A – agent-to-agent

    Your agent delegates tasks to agents on other platforms – including ones running on-premises.5

  • more than 1,000

    Connectors

    Prebuilt connectors to SAP, Salesforce, ServiceNow and many more, plus your own custom connectors.6

  • no API needed

    Computer use

    The agent operates websites and desktop apps like a person – for legacy systems without an interface.7

  • third party

    External agents and models

    Connected agents, for example from Microsoft Foundry or Fabric (preview), and models from Anthropic or Mistral.8

Where Copilot Studio isn't the best choice, we say so – and show you the alternative.

Autonomous – under control

How do you stay in control of autonomous agents?

Autonomous agents are started by events and work in the background. You decide what they may decide on their own: routine flows through, exceptions go to people with full context. Every step is logged, every agent has its own identity.

An order, without anyone asking

Sample run, simplified

  1. 6:02 a.m.Trigger

    An order arrives as a PDF in the shared mailbox.

  2. 6:02 a.m.Understand

    The agent reads the document: customer, 14 line items, requested date.

  3. 6:03 a.m.Check

    Master data, terms and delivery date matched in the ERP.

  4. 6:03 a.m.Mismatch

    Line 7 is 8% below list price. The agent asks sales operations for approval.

  5. 7:41 a.m.Human

    Sales operations approves the special price – one click in Teams.

  6. 7:42 a.m.Done

    Order created in the ERP, confirmation sent to the customer, case logged.

Autonomy is a dial, not a switch

You set the dial for each process step – not for the agent as a whole.

  1. SuggestThe agent prepares, a person executes.
  2. After approvalThe agent acts once a person approves.
  3. Within a thresholdThe agent acts on its own up to a defined limit, above it only with approval.
  4. With escalationThe agent acts independently and escalates exceptions and uncertainty.

Human-in-the-loop

The agent brings in a person at defined points – wherever responsibility sits.

Approvals

Multi-stage approvals in workflows. Routine cases can be pre-checked by AI, anything critical is always decided by a person.

Requests for information

The agent pauses, asks the responsible person for missing information via Outlook and then carries on.9

Handover with context

In customer service, your team takes over with the full history – nobody has to repeat anything.

Audit log

Every run is traceable: trigger, tools, decisions.

Identity and permissions

Since July 2026, every new agent automatically receives its own Microsoft Entra Agent ID – opting out is no longer possible. Access rights, Conditional Access and DLP apply per agent.10

Budgets

Credit allocations per environment plus the Agent Usage Estimator (standard harness) and the Copilot Credit Estimator (GitHub Copilot harness) make costs predictable.11

Evidence

The market is moving – and many stumble at the start

Agentic AI is moving into enterprise applications. At the same time, many projects fail on cost, unclear value or missing controls. Four numbers that show both:

40%

of enterprise applications will feature task-specific AI agents by 2026 – up from less than 5% in 2025.

Gartner, August 26, 202512

over 40%

of agentic AI projects will be canceled by the end of 2027 – due to escalating costs, unclear business value or inadequate risk controls.

Gartner, June 25, 202513

at least 15%

of day-to-day work decisions will be made autonomously through agentic AI by 2028 – up from 0% in 2024.

Gartner, June 25, 202513

106–314%

projected three-year ROI for a composite organization using Copilot Studio (low to high scenarios). Forrester TEI study commissioned by Microsoft.

Forrester Consulting, study commissioned by Microsoft14

That's why we start with the business case, a baseline and governance – not with building.

What's new in 2026

What's new in Copilot Studio in 2026?

As of September 2026

Since August 2026, Copilot Studio handles multi-step processes with files, skills and memory – the chatbot builder has become a platform for process agents.

  1. Copilot Studio extension for Visual Studio Code10

    GA

  2. Claude models available globally; agent evaluations including multi-turn tests; Work IQ tools (preview)10

    GA

  3. Agent-to-agent protocol (A2A); Agent Usage Estimator for credit forecasts; real-time voice agents (preview)10

    GA

  4. Computer use: agents operate web and desktop apps without an API; automatic Microsoft Entra Agent IDs (preview)10

    GA

  5. Microsoft Agent 365: one control plane for all agents15

    GA

  6. Claude Sonnet 5 and GPT-5.5 Chat generally available; GitHub Copilot harness in production-ready preview with skills and memory (memory in preview)10

    GA

  7. Copilot Cowork in Microsoft 365: an agentic system for complex, long-running tasks16

    GA

  8. Automatic Microsoft Entra Agent ID for every new agent (no opt-out); MCP server certification (preview); real-time agents also on digital messaging (preview)10

    GA

  9. GitHub Copilot harness generally available: multi-step processes with files, skills and memory (memory in preview)17

    GA
Technical FAQ

Questions from IT and architecture

Seven questions that come up in technical conversations – with dates and sources.

What is Microsoft Copilot Studio – and how is it different from Microsoft 365 Copilot?

Microsoft 365 Copilot supports individuals in Word, Outlook and Teams; Copilot Studio is the platform for building and running AI agents for entire processes.

Copilot Studio adds process logic, integrations with your systems, your own channels such as website or phone, a free choice of models and governance. Agents built in Copilot Studio can also be used in Microsoft 365 Copilot and Teams.

Copilot, Copilot Chat, Cowork, Agent Builder, Copilot Studio – what's what?

Copilot Chat is the chat, Microsoft 365 Copilot is the AI in your Office apps, Cowork is the agentic system for longer tasks, Agent Builder is the entry point for simple agents – and Copilot Studio is the platform for process agents.

The Copilot terms at a glance
TermWhat it is
Microsoft 365 Copilot ChatWeb-based chat, included in many Microsoft 365 licenses.
Microsoft 365 CopilotPer-user license: AI in Word, Excel, Outlook and Teams with your work context.
Copilot CoworkAgentic system in Microsoft 365 for long-running, multi-step tasks, generally available since June 16, 2026.16
Agent BuilderSimple agents for end users, right inside Microsoft 365 Copilot.
Copilot StudioPlatform for agents with process logic, integrations, your own channels, model choice and governance. Agents from Microsoft 365 Copilot can be moved over.
Microsoft FoundryPro-code platform for custom AI solutions and models.18
What is an AI agent – and how is it different from a chatbot or RPA?

An AI agent pursues a goal, plans the steps itself and acts in your systems through tools.

A classic chatbot follows predefined dialogs, RPA replays fixed click paths. An agent understands unstructured input such as emails or PDFs, decides within guardrails and brings in a person for exceptions. Where a flow is fixed and stable, a classic workflow is often still the better choice.

Which AI models can I use in Copilot Studio?

As of September 2026: models from OpenAI (GPT), Anthropic (Claude) and Mistral, plus more than 1,800 models from Microsoft Foundry for individual prompt steps.

On the GitHub Copilot harness, the options include GPT-5.5 Chat, GPT-6 Astra, Claude Sonnet 5 and Claude Opus 5. External models (Anthropic, Mistral) must be enabled by an admin in both the Power Platform admin center and the Microsoft 365 admin center. Anthropic operates as a Microsoft subprocessor. In Europe, Claude models are cross-geo and may process data outside the region. The Auto category is currently available only as a preview on the standard harness. The current overview is in the model table above.2

What is a harness – and which one do we need?

The harness is the runtime between agent and model; which one fits depends on the process.

The GitHub Copilot harness is built for multi-step processes with files, skills and memory, the standard harness for rule-based, predictable flows, the Copilot chat harness for knowledge in Microsoft 365 Copilot Chat. The GitHub Copilot harness has been generally available since August 3, 2026. The GitHub Copilot harness is not the GitHub Copilot service. Customer data is not sent to GitHub Copilot; Copilot Studio's privacy and data residency commitments continue to apply.1

How do we stay in control of autonomous agents?

With clear guardrails: an identity for every agent, least-privilege permissions, approvals at critical points and a complete audit log.

Since July 2026, every new agent automatically receives its own Microsoft Entra Agent ID – opting out is no longer possible.10 Microsoft Agent 365, generally available since May 1, 2026, brings inventory, policies, security and lifecycle of all agents together – including partner agents; list price USD 15 per user per month, or included in Microsoft 365 E7.15 Add evaluations with test sets and human-in-the-loop wherever responsibility sits.

What happens when a model is replaced?

Models are replaced all the time – a good agent survives that because model changes are planned and tested.

For example, GPT-4o was replaced as the default model by GPT-4.1 in October 202510; today the default is GPT-5.5 Chat. Since August 2026, Microsoft Learn has had a dedicated guide on the model lifecycle.19 We safeguard every agent with evaluations, including multi-turn tests, and check new models before they go into production.

Glossary

Key terms explained

AI agent
Software that pursues a goal, plans the steps itself and acts in systems through tools – within defined guardrails.
Autonomous agent
An AI agent started by an event such as an incoming email or a new record, working without a prompt. Exceptions go to people.
Harness
The runtime between agent and language model: it manages context, model calls and tools. Copilot Studio has three harnesses: the GitHub Copilot harness, the standard harness and the Copilot chat harness.
Copilot Credit
The billing unit of Copilot Studio. Each feature consumes a defined number of credits; list price USD 0.01 per credit (as of September 2026).
MCP
Model Context Protocol: an open standard that lets systems describe their tools and data to AI agents.
A2A
Agent-to-agent protocol: an open standard that lets agents on different platforms exchange tasks.
Human-in-the-loop
The agent brings in a person at defined points – for approvals, requests for information or exceptions.
Computer use
A tool that lets an agent operate websites and desktop apps with mouse and keyboard – for systems without an interface.
Grounding
Anchoring an agent's answers in approved sources such as SharePoint or Dataverse, with citations.
Agent 365
Microsoft's central control plane for all agents in an organization: inventory, policies, security and lifecycle, including partner agents.

Technical questions are best answered with a concrete process. Your IT team is welcome to join.

Book a call with IT
Sources

Sources and date of information

All information as of September 24, 2026. Prices are list prices in USD and depend on contract and purchasing channel. Features in preview or experimental status are not intended for production use.

  1. Harnesses in Copilot Studio · Microsoft Learn · retrieved September 24, 2026
  2. Model availability and controls (GitHub Copilot harness) · Microsoft Learn · retrieved September 24, 2026
  3. Select a primary AI model for your agent · Microsoft Learn · retrieved September 24, 2026
  4. Use agent tools to extend, automate, and enhance your agents · Microsoft Learn · retrieved September 24, 2026
  5. Connect an agent available over the Agent2Agent (A2A) protocol · Microsoft Learn · retrieved September 24, 2026
  6. Custom connectors overview · Microsoft Learn · retrieved September 24, 2026
  7. Automate web and desktop apps with computer use · Microsoft Learn · retrieved September 24, 2026
  8. Add other agents overview · Microsoft Learn · retrieved September 24, 2026
  9. Request information from human review in agent flows · Microsoft Learn · retrieved September 24, 2026
  10. What's new in Copilot Studio · Microsoft Learn · retrieved September 24, 2026
  11. Govern Copilot Credit consumption for agents powered by the GitHub Copilot harness · Microsoft Learn · retrieved September 24, 2026
  12. Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025 · Gartner · August 26, 2025
  13. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 · Gartner · June 25, 2025
  14. New Technology: The Projected Total Economic Impact™ Of Microsoft Copilot Studio · Forrester Consulting · study commissioned by Microsoft
  15. Microsoft Agent 365, now generally available, expands capabilities and integrations · Microsoft Security Blog · May 1, 2026
  16. Copilot Cowork is now generally available · Microsoft AI at Work Blog · June 16, 2026
  17. More powerful agents and workflows for autonomous business processes: Introducing a new harness for Copilot Studio · Microsoft Tech Community, Copilot Studio Blog · August 3, 2026
  18. Choose the right platform · Microsoft Learn · retrieved September 24, 2026
  19. Manage the AI model lifecycle for Copilot Studio agents · Microsoft Learn · retrieved September 24, 2026