Knowledge for better technology decisions
Knowledge BaseAI FinOps

AI FinOps · Copilot Studio

Three Copilot Studio harnesses:
Which one fits which agent?

In Copilot Studio, the model is not the only thing that shapes an agent. The harness controls how a goal is processed, how knowledge and tools are used, how failures are handled, and how usage is billed. Choosing a harness is therefore an architecture, governance, and cost decision.

The shape of the work matters more than the longest feature list.

A clearly defined conversation does not need a reasoning-heavy runtime. A complex end-to-end process, however, cannot be reduced reliably to rigid conversation paths.

  1. 01Fixed topics, rules, and predictable paths: standard harness
  2. 02Enterprise knowledge directly in Microsoft 365 Copilot Chat: Copilot chat harness
  3. 03Multistep goals, tools, and adaptive execution: GitHub Copilot harness

One platform – three runtime logics

The harness shapes
how the agent works.

Microsoft describes the harness as the runtime layer between the agent design and the model. It determines when the model is called, which components provide context, how responses are interpreted, and when tools are invoked.

01

Not the model, not the interface, and not the licence

These layers are often mixed together. Separate them before making an architecture and cost decision.

01

Agent design

The goal, instructions, knowledge, tools, identity, and permitted actions describe what the agent is allowed to do.

02

Harness

The runtime orchestrates context, model calls, tools, intermediate steps, and responses to unexpected results.

03

Model

The language or reasoning model processes the context it receives. It is one component inside the runtime.

Important:

A GitHub Copilot harness agent is not the same as a GitHub Copilot user licence. Here, the term refers to the Copilot Studio runtime and its billing model.

02

Three different answers to three shapes of work

Start with the required outcome. Then assess how predictable the path is and how much planning, tool use, and recovery the agent must handle.

01PREDICTABLE

Standard harness

Best suited for
Rule-based agents, defined conversations, FAQs, help desks, and agent flows.
How it works
Makers author topics, rules, prompts, and paths. The agent follows the designed structure.
Example
An IT help desk answers known questions and routes clearly defined exceptions.
Important boundary
Open-ended goals and dynamic replanning are not the primary focus.
Licence & cost
Standard licence and credit rules; qualifying Microsoft 365 Copilot usage may be included.
02IN M365 CHAT

Copilot chat harness

Best suited for
Focused internal assistants that extend Microsoft 365 Copilot Chat with enterprise knowledge.
How it works
The agent stays inside the Copilot Chat experience and answers specialised questions from approved sources.
Example
An onboarding agent explains policies and roles using SharePoint content.
Important boundary
Internal teams only; not intended as an external customer channel.
Licence & cost
Included with Microsoft 365 Copilot or consumption-based, depending on the scenario.
03AGENTIC

GitHub Copilot harness

Best suited for
Reasoning-heavy, multistep work and complex end-to-end business processes.
How it works
The agent decomposes goals, calls knowledge, connectors, MCP servers, or connected agents, and adapts its steps.
Example
An accounts-payable agent reads invoices, matches purchase orders, and routes exceptions.
Important boundary
Greater autonomy requires stronger evaluation, safeguards, monitoring, and budget controls.
Licence & cost
Always usage-based; credits can accrue during creation, preview, testing, and evaluation.
03

What changes technically and operationally?

The differences extend from orchestration to publishing. This table highlights the main design, operations, and governance implications.

FeatureStandard harnessCopilot chat harnessGitHub Copilot harness
Best forRule-based conversations and agent flowsSpecialised knowledge agents in Microsoft 365 CopilotComplex, multistep business processes
ControlDefined topics, rules, and pathsEnterprise grounding inside Copilot ChatGoal-led planning and dynamic tool selection
RecoveryDesigned into the authored pathLimited to the focused chat taskCan adapt steps and try alternative approaches
FilesThrough configured answers and actionsWithin the Microsoft 365 Copilot contextCan create or edit Word, Excel, PowerPoint, and PDF files
Skills & memoryNot the primary operating modelFocused on chat extension and knowledgeAvailable as dedicated components
PublishingInternal and external channelsInternal Microsoft 365 users onlyInternal teams or external customers
BillingStandard licence and Copilot Credit rulesIncluded or consumption-based, depending on scenarioAlways usage-based in Copilot Credits
04

The same word ‘agent’ – three suitable architectures

The actual process usually reveals the right harness faster than a feature list.

STANDARD

IT help desk with fixed rules

The agent recognises known requests, guides users through defined steps, and routes exceptions to the service desk.

Priority: predictability and controlled paths
COPILOT CHAT

Employee onboarding knowledge

The agent answers policy, role, and process questions inside Microsoft 365 Copilot using approved enterprise knowledge.

Priority: knowledge in the existing work context
GITHUB COPILOT

Invoice to approval

The agent reads documents, matches purchase orders, calls systems, and routes exceptions across dependent steps.

Priority: planning, tools, and adaptation
05

The harness changes when and how consumption occurs

For FinOps, the agent licence alone is not enough. Runtime, user, identity, channel, trigger, model, and tool calls all matter.

STANDARD

Scenario-dependent

Standard-harness agents use Copilot Credits at published feature rates. Qualifying authenticated business-to-employee usage can be included for Microsoft 365 Copilot licensed users; other scenarios consume credits.

COPILOT CHAT

Check user and context

Usage can be covered by a Microsoft 365 Copilot user licence or billed by consumption. User licence, identity, Microsoft 365 channel, and invoked capabilities determine the outcome.

GITHUB COPILOT

Consumption during build

Credits can accrue during natural-language creation, preview, testing, evaluation, and production. Billing can include model tokens, tools, and harness usage.

FinOps rule:

Plan separate build, test, and production budgets, set per-agent limits, and measure cost per successfully completed business outcome—not cost per prompt.

06

Five questions before choosing a harness

Document the answers before building. Microsoft states that an agent cannot be transferred between the standard and GitHub Copilot harnesses.

  1. 01

    Is the route to the outcome predictable?

    If rules and paths are clear, the standard harness is often the more economical and controllable choice.

  2. 02

    Should the agent live directly in Microsoft 365 Copilot Chat?

    For a focused internal knowledge assistant, the Copilot chat harness is the natural option.

  3. 03

    Must the agent plan, combine tools, or recover when something fails?

    The GitHub Copilot harness is a better fit for that pattern, with greater testing and cost-control needs.

  4. 04

    Who will use the agent and through which channel?

    Internal or external users, authentication, and publishing destination affect both harness fit and billing.

  5. 05

    How will risk, credits, and value be governed?

    Ownership, data access, approvals, limits, monitoring, and success metrics belong in the architecture decision.

As deterministic as possible. As agentic as necessary.

Do not default to the most powerful runtime. Choose the simplest harness that reliably delivers the outcome, and design cost controls, safeguards, and measurement from the start.

Verified against Microsoft sources.

This guidance is based on official Microsoft information available on 13 August 2026. Features, preview status, and billing rules may change.

Note: This article provides professional orientation and does not replace a review of your specific licence terms, tenant configuration, Product Terms, or contract documentation.

Deep dive · Copilot Credits

When do Copilot Studio agents consume Copilot Credits?

Evaluate user, channel, identity, trigger, and harness together.