A New Harness for Copilot Studio: What Has Changed
On copilotstudio.microsoft.com, a Try now button now offers the ability to switch to a completely redesigned authoring experience. This is not just a simple interface redesign: under the hood, it's a new orchestrator — often called the new harness — that drives the construction and execution of agents. It incorporates the logic already present in Copilot in agentic mode (the famous "tenacious" behavior, which persists until the task is resolved rather than abandoning after the first failure).
This change is significant for an administrator or solution architect:
- The old authoring experience (multiple tabs, subjects, topics) remains accessible, but becomes a parallel path rather than the default route.
- The new canvas is intentionally streamlined: a single screen groups together model, skills, tools, knowledge, memory, and instructions.
- It is possible to return to the old experience at any time via a dedicated link, simply by indicating a reason for returning.
Coexistence of Two Experiences
Agents already built with the old experience continue to work. The choice between old and new harness is made at the creation of each agent, not at the tenant level.
Choosing the Right AI Model (and Managing Credit Consumption)
The first decision to make on a new agent is the choice of language model. The list offered depends on your tenant and the models enabled by your administrator — it typically includes models from the GPT-5 family as well as Claude models (Sonnet and Opus variants). This choice is not cosmetic: each call to the model consumes Copilot credits, and the most capable models (often those enabled by default, such as "Opus" variants or extended reasoning modes) are also the most expensive in terms of credits.
It is this same model that explains the famous "tenacity" observed on Copilot agents in agentic mode: where an old Copilot Studio agent would abandon after a failure with a message like "I wasn't able to do it," the new harness retries, adjusts its approach, and completes the requested task — at the cost, of course, of higher credit consumption in case of multiple iterations.
Licensing and Billing Impact
Unlike the old experience, where some usage could be covered by a Microsoft 365 Copilot license, the new harness systematically bills in Copilot credits — including during testing and preview phases. Check your credit pool before multiplying preview sessions.
To quickly verify assigned licenses before widely deploying these agents, a check via Microsoft Graph is useful:
1# Prerequisites: Microsoft.Graph module (Connect-MgGraph -Scopes "User.Read.All")2Get-MgUserLicenseDetail -UserId "user@yourdomain.com" |3 Select-Object SkuPartNumber, SkuIdThis allows you to confirm, before deploying an agent in production, which users actually have a license covering the use of Copilot Studio and associated credits.
Skills: The Standard That Unifies Copilot Studio and Copilot
The real structural innovation is the arrival of skills in Copilot Studio — the same format used by Copilot in agentic mode, by Claude, or by Open AI compatible tools. A skill is based on a skill.md file: a reusable recipe that describes a series of steps to execute systematically (for example, applying a brand style to every generated PowerPoint presentation).
Since this format has become a de facto standard, a skill created in Copilot (tab Customize > Skills) is directly portable to Copilot Studio. The flow is straightforward:
Locate the Skill Folder
In Copilot, skills are stored locally under OneDrive > Documents > Copilot > skills. Each skill has its own subfolder containing skill.md and, possibly, ancillary files (logos, templates, blueprints).
Compress the Complete Folder
Simple download from the Copilot interface only retrieves the skill.md file, not the associated resources. Therefore, you must manually compress the entire folder:
1# Compresses all folder content (skill.md + resources)2Compress-Archive -Path "C:\Users\you\OneDrive\Documents\Copilot\skills\brand-guide\*" `3 -DestinationPath "C:\Users\you\Desktop\brand-guide.zip"Import the Skill into Copilot Studio
In the agent, Skills section, click Upload and select the brand-guide.zip file. After a few seconds, the skill appears with all its instructions and resources.
If you start from scratch, you can create a blank skill (Create from blank) and write the instructions yourself. However, manually writing 150 to 200 lines of instructions is tedious: the most efficient method is to ask Copilot to ask you the necessary questions (brand colors, fonts, logos…) and then generate the corresponding skill.md file itself.
A Skill Doesn't Have to Be New
Any recipe already used in an old topic or recurring business process can be reformulated as a skill. The goal: capitalize on what already works rather than rebuild everything in the new formalism.
Tools, MCP, and Knowledge Sources: The Mechanics Under the Hood
Alongside skills, we find tools — the classic connectors (SharePoint, Dataverse, MCP servers, etc.) already present in Power Apps, Power Automate, and the old Copilot Studio. Adding a tool like Get items on SharePoint requires configuring:
- The name and description of the tool: this is what the orchestrator uses to decide when to invoke it. A generic name ("Get items") should be renamed explicitly ("Get inventory data") so the agent knows when to actually use it.
- The authentication mode: on behalf of the end user or on behalf of the agent creator — a choice that directly impacts the security and traceability of access.
- The input parameters (SharePoint site, list name, OData filtering query): each field can be a fixed value or dynamically generated by the AI from a natural language description.
A Poorly Named Tool Breaks the Entire Agent
If the agent is not using the right tool at the right moment, the most frequent cause is an imprecise name or description on the tool — not an orchestrator bug. Before debugging the agent's instructions, systematically verify this point first.
On the knowledge side, the agent can query the web via Bing by default (a disableable parameter), specific public sites, SharePoint, OneDrive for Business, or directly uploaded files.
| Knowledge Source | Permission Respect | Typical Use Case |
|---|---|---|
| SharePoint / OneDrive for Business | Yes — responses filtered according to end user rights | Sensitive business data, HR or financial documents |
| File Uploaded Directly to the Agent | No — accessible to all agent users | Public or generic documentation without sensitive data |
| Public Websites | Not applicable | Monitoring, product FAQ, public marketing content |
If a tenant-wide DLP (Data Loss Prevention) restricts certain connectors, it is useful to check the status of policies before multiplying tools on your agents:
1# Required module: Microsoft.PowerApps.Administration.PowerShell2# Minimum role: Power Platform Administrator3Add-PowerAppsAccount4Get-AdminDlpPolicy | Select-Object DisplayName, EnvironmentType, CreatedTimeMemory, Child Agents, and Data Security
Two capabilities still in preview deserve particular mention. Memory allows an agent to retain information provided by a user over the course of interactions (for example, their assigned region), so it doesn't have to ask for it again at each session. This memory is specific to the agent — it is not shared between different agents.
The agent-to-agent call allows you to attach an existing agent as a sub-component of a parent agent, to build modular solutions without duplicating the business logic already encapsulated elsewhere.
Preview Features
Memory and evaluation remain in preview at the time of writing. They already work convincingly in exploratory tests, but should not yet support critical scenarios in production until they reach general availability.
Instructions, Preview Mode, and Usage-Based Billing
Instructions define the agent's mission ("your role is to generate PowerPoint presentations; use the brand-guide skill for any formatting request") and, possibly, its personality. They remain the most determining variable for response quality — more so than the choice of model itself in many cases.
Preview mode allows you to test the agent under real conditions, with two display levels:
- The detailed view (default), which exposes each tool call, each reasoning step — useful for debugging.
- The End user preview button, which hides all technical details to provide a faithful preview of what the end user will see.
Each Test Consumes Tenant-Wide Credits
Unlike the old experience, where the design phase was not billed, each preview execution — generation, tool call, skill call — consumes Copilot credits charged to the tenant. A repeated test on an expensive model (such as Opus) can quickly impact the allocated budget.
To check consumption after several test sessions, the agent's Monitor tab (accessible after first save) provides a view of sessions and associated billing. Remember to click Save regularly: there is no automatic saving in this new interface.
Publishing Your Agent: Teams, M365 Copilot, and Web Applications
Once the agent is validated, the Publish menu offers several distribution channels:
- A demo website (unavailable if the agent uses authentication specific to a connector).
- Web application integration via a code snippet to embed.
- Availability in Teams and Microsoft 365 Copilot, the most common channel for internal organization distribution.
Publication typically takes one to two minutes to propagate. Once active, the channels (Teams, M365 Copilot) appear as "enabled" in the Availability options tab, and the agent becomes immediately available via Add an agent in Microsoft 365 Copilot.
Key Takeaways
- The new Copilot Studio harness is based on the same orchestrator as Copilot in agentic mode: more tenacity, but systematic billing in credits, including during the testing phase.
- Skills (
skill.mdformat) are now interoperable between Copilot and Copilot Studio — a real leverage for reuse in recurring business processes (branding, reporting, document formatting). - The name and description of tools are the first control point to verify when an agent is not using the right connector at the right time.
- Direct file upload to an agent bypasses native permissions (SharePoint, OneDrive) — reserve this for non-sensitive content.
- Memory and child agents open perspectives for modular solutions, but remain in preview: validate before any critical deployment.
- Building Copilot Studio agents for repetitive tasks allows removing these uses from the individual Copilot user credit pool and attaching them to a controlled budget on the organization side — a concrete lead for limiting the cost of adopting generative AI at scale.
Before generalizing this approach across all your teams, test it on a circumscribed use case (one skill, one tool, one publication channel) to precisely assess the impact in credits before opening access to a larger audience.
