Compare Copilot Studio agent harnesses

Completed

Copilot Studio provides three harnesses for different types of work. To select one, consider what the agent must accomplish, how much control you need, and where people access it.

Use the Microsoft Copilot harness for reasoning-heavy work

The Microsoft Copilot harness supports agents and workflows that complete complex, multistep work. It can interpret a goal, create a plan, call multiple tools, observe results, and adjust when a request changes or an action fails.

This harness is appropriate when the solution needs to:

  • Reason through ambiguous decisions.
  • Coordinate several tools, data sources, or agents.
  • Complete longer business processes that can't be fully scripted in advance.
  • Use features such as skills and memory.

You guide these agents with natural-language instructions, knowledge, tools, skills, and boundaries. You don't author every possible conversation route. Agents and workflows powered by this harness use Copilot Credits for usage-based billing.

For example, an accounts-payable agent can read an invoice, locate its purchase order, compare the values, request missing evidence, and route an exception. The correct path depends on the content of each invoice and the results returned by business systems.

Use the standard harness for predictable paths

The standard harness supports rule-based agents and agent flows. It is appropriate when you need structured, repeatable conversations or processes with predictable behavior.

It supports both classic and generative orchestration. With generative orchestration, a standard-harness agent can dynamically select topics, tools, knowledge, and connected agents. Reasoning or using multiple tools isn't exclusive to the Microsoft Copilot harness. Choose based on the complete capability set, authoring requirements, and existing investments. For details, see Orchestrate agent behavior with generative AI.

This harness is appropriate when the solution needs to:

  • Follow topics and paths that a maker defines.
  • Use specific trigger phrases, questions, conditions, and branches.
  • Provide tightly controlled responses for well-understood requests.
  • Reuse existing standard-harness topics, prompts, or agent flows.

For example, an IT team can extend an existing equipment-return agent by reusing its standard-harness topics and agent flows. Those topics collect an asset number, confirm the employee's location, and route the conversation through approved shipping options. The standard harness remains fully supported alongside the Microsoft Copilot harness.

Use the Customer Conversation harness for customer service

The Customer Conversation harness supports agents that answer customer questions and resolve issues over voice and chat. It combines predictable responses for defined topics with natural-language conversation. Its multi-agent architecture uses specialized sub-agents, each with a defined escalation path to a human representative.

This harness is appropriate when the main requirement is to:

  • Support customers through website chat or a phone number.
  • Delegate different customer intents to specialist sub-agents.
  • Use knowledge, tools, skills, and memory during a conversation.
  • Escalate unresolved issues to a human representative through a defined path.

For example, a retailer needs a customer service agent for order questions over website chat and phone calls. Specialist sub-agents handle delivery and returns, while unresolved disputes transfer to a human representative. The main requirement is customer conversation and escalation, not a long-running file-processing task.

Agents on this harness use Copilot Credits for usage-based billing. The Customer Conversation experience is in preview. For capabilities and availability, see Agents powered by the Customer Conversation harness overview.

Compare the harnesses

Aspect Microsoft Copilot harness Customer Conversation harness Standard harness
Primary use Reasoning-heavy, multistep processes Customer service over voice and chat Structured conversations and existing topic-based solutions
Main design approach Goals, instructions, knowledge, tools, and skills Specialist sub-agents with defined human escalation paths Topics and tools with classic or generative orchestration
Recovery Can adjust its plan based on results Escalates unresolved issues to a human representative Uses configured topic paths and orchestration behavior
File-intensive work Supports creating, editing, and reasoning over files Not a primary focus Not a primary focus
Skills and memory Supported Supported Not a primary focus
Publishing Internal teams or external customers Website chat or a phone number Internal teams or external customers
Billing Copilot Credits under harness-specific usage rates Copilot Credits under customer-conversation usage rates Copilot Credits under standard-harness billing rules

Skills and memory have different implementations across harnesses. Customer Conversation supports both instructions-based Generative skills and deterministic Topic skills with authored steps and conditions. Its memory captures structured customer context, while Microsoft Copilot harness memory retains relevant details from interactions. Don't assume that a shared feature name means identical configuration or behavior. See Skills overview and Memory for Customer Conversation.

Note

Copilot Credits are used across harnesses, but metering and entitlements differ. For the Microsoft Copilot harness, building, previewing, testing, and creating and running evaluations can consume credits before publication. Standard-harness billing starts after publication. For current rules, see Overview of usage-based billing and Licensing for agents powered by the standard harness.

Apply a selection sequence

Use these questions when selecting a harness:

  1. Is the primary job customer service over voice and chat, a reasoning-heavy business process, or a rule-based conversation?
  2. Does the solution need specialist sub-agents and human escalation, explicit maker-defined paths, or dynamic planning toward a goal?
  3. Does the work require multiple tools, memory, or recovery from intermediate failures?
  4. Where must users access the completed agent?
  5. Which licensing and consumption model fits the organization?

These questions help separate a solution's actual requirements from the appeal of any individual feature. Capabilities overlap, so a need for reasoning, fixed steps, or tools alone doesn't uniquely identify a harness.

For the current comparison, see Harnesses in Copilot Studio.

Note

We recognize that different people like to learn in different ways. You can choose to complete this module in video-based format or you can read the content as text and images. The text contains greater detail than the videos, so in some cases you might want to refer to it as supplemental material to the video presentation.