Copilot Studio runs on three harnesses. Two were already operating before August. The third arrived on 3 August, and an agent created on any of them cannot be moved to another.
Microsoft defines the harness as the runtime that sits between what you design and the model that reasons: it decides when to call the model, what it sends and which tools it uses. It is not the model, and it determines what you can build. Producing files without templates, using skills or remembering across sessions depends on the harness, not on the model you pick.
Four findings
- 01
Two were already operating before they had a name, and they are not decision trees
The standard harness powers most agents built so far, and since 2026 it ships with generative orchestration: the agent picks topics and tools by their descriptions. The Copilot chat harness is the one behind Microsoft 365 Copilot Chat. «Harness» is new nomenclature, not a new engine.
- 02
The choice is irreversible
The documentation is literal: an agent created with the GitHub Copilot harness cannot be transferred to the standard one, nor the other way round. You choose at creation, and you live with it.
- 03
Only one of the three bills from build time
The standard harness starts billing after you publish. The GitHub Copilot harness charges credits from the moment you start building, including the test panel and the generation of evaluations.
- 04
The official table says «no» where the answer is «yes, in another way»
The standard harness has skills, produces documents and retains state. None of them is the same as in the new harness, but neither is it the absence the table suggests.
The three, side by side
| Criterion | GitHub Copilot harness | Standard harness | Copilot chat harness |
|---|---|---|---|
| What it is for | Complex, multi-step business processes | Rule-based agents and structured conversations | Extending Microsoft 365 Copilot Chat |
| How it works | Reasons about a goal on its own, step by step | Works with the topics and tools you define | Connects knowledge to Copilot Chat |
| Orchestration | Fixed, imposed by the harness | Configurable: generative or classic, by trigger phrases | Not applicable |
| On error | Retries and looks for alternative routes | Follows the routes you built | Not its focus |
| Files | Reads them and also creates and edits them natively: Word, Excel, PowerPoint and PDF | Reads them as knowledge. Generates Word through Document output, with a template and a flow | Reads them via SharePoint and Graph connectors |
| Skills | A SKILL.md file in Markdown, no code | Azure Bot Service skills: a pro-code bot called from a topic | Not applicable |
| Memory | Persistent per user, across sessions. Cleared after 28 days without use | Topic and global variables, within the conversation | Not applicable |
| Publishing | Internal teams or external customers | Internal teams or external customers | Internal teams only |
| Billing | Credits, from build time | Licensing, from publication | Consumption, or included in the licence |
| Migration | Does not exist in any direction | Does not exist in any direction | Does not exist in any direction |
The same-named capabilities, and how they actually differ
- 01
Skills
In the new harness it is a SKILL.md file: YAML front matter with name and description, instructions in Markdown, packageable as a ZIP with scripts and templates. In the standard harness it is a bot deployed with the Bot Framework SDK or the Microsoft 365 Agents SDK, called from a topic. One is a text file, the other is a service.
- 02
Memory
It is switched on per agent and stores a separate memory per user, managed by Microsoft. It is cleared after 28 days without interaction and is disabled in group chats and Teams channels.
- 03
Files
The difference is not producing, it is producing without scaffolding. Microsoft writes that this harness «natively creates and edits» Word, Excel, PowerPoint and PDF. The standard harness also produces a document, but by another route: a Word template with fields to replace, a prompt that fills them and a flow that saves the file.
What caught my attention was not the new harness. It was the sentence saying there is no migration between harnesses. On a low-code platform the implicit promise was always that rebuilding is cheap. Not here: the first agent someone on your team creates this week is tied to an engine and to a billing model, and the only way out is building it again.
This report reflects the documentation in force on the date shown. These conditions change often: check the current version before making a decision.