What inline agents change in Copilot Studio
Microsoft Copilot Studio has significantly overhaul its creation experience. The old concept of standalone "prompts" has disappeared in favor of two distinct objects: agents and workflows. These two types of resources rely on the same development environment as GitHub Copilot, and can also be built from Microsoft 365's Agent Builder.
The key novelty is the inline agent: an agent that exists exclusively within the context of a workflow. It is not registered in the Copilot Studio catalog, does not appear on the home page, and is not reusable outside the workflow that contains it. This is a deliberate design choice — when the logic is too specific to be shared, there's no point in cluttering the catalog.
License Prerequisites
Using the Researcher agent and displaying conversations in M365 Copilot require an active Microsoft 365 Copilot license. Copilot Studio workflows fall under the standard Copilot Studio license.
Three types of agents available in a workflow
When adding an agent node to a workflow, three options are available:
- New agent for this workflow: inline agent, invisible outside the workflow.
- Existing published agent: allows reusing an agent already deployed in Copilot Studio — useful for repeatable skills or agents shared across multiple workflows.
- M365 Copilot Agent: leverages native agents from Microsoft 365 Copilot, notably Researcher and Analyst.
Unpublished Agents
Only published agents can be referenced from a workflow. Draft or unpublished agents are not available in the selection list.
Automate monitoring with Researcher and workflows
The Researcher agent, natively available in Microsoft 365 Copilot, is particularly suited for automating monitoring reports. The principle is simple: trigger a workflow on a schedule or business event, inject a topic via dynamic content, and let Researcher produce a structured report.
The prompt template provided by Microsoft for Researcher begins like this:
Create a comprehensive, well-structured report on the current state and future trajectory of [topic]
In a workflow, this prompt is enriched with dynamic content — a topic variable from a manual trigger, Teams message, email, or file dropped in a SharePoint library.
Create a workflow with manual trigger
In Copilot Studio, create a new workflow. Select the Manual trigger and add a text input named topic with the label "Please enter your topic".
Add an M365 Copilot agent node (Researcher)
Add an agent node and select M365 Copilot > Researcher. In the message field, paste the search prompt and inject the dynamic variable topic at the subject location.
1Create a comprehensive, well-structured report on the current state 2and future trajectory of [topic]. Do not ask me extra questions. 3About one to three pages.Enable display in M365 Copilot
In the node settings, enable the Show conversation in M365 Copilot option. This allows you to find the result directly in the Copilot interface, without having to consult the workflow history.
Publish and test
Publish the workflow, then launch a manual execution by entering a test topic — for example: Inline agents in workflows of the new Copilot Studio. The result appears in the M365 Copilot conversation a few moments later.
The operational benefit is concrete: by setting up a nightly recurrence, IT teams find a monitoring report already produced in their Copilot interface the next morning, without manual intervention.
Structured output: goodbye to manual JSON parsing
One of the most significant improvements for integrators is the structured output of inline agents. Until now, retrieving an exploitable response from an agent in Power Automate involved working with raw JSON and using Parse JSON actions — a frequent source of errors and fragility.
Structured output allows you to directly define the expected properties, with their type:
| Property | Type | Example value |
|---|---|---|
| name | Text | Lost key card - Andrew Hess |
| severity | Choice: low / medium / high | high |
| owner | Text | Security team |
| description | Text | Key lost in a public area |
These properties are then directly accessible as dynamic content in the following workflow steps — to write to a SharePoint list, SQL table, Dataverse, or Excel file.
Use case: ticket processing
An incoming ticket (from ServiceNow, a form, or an email) is sent to an inline agent configured with GPT-5 chat. The agent analyzes the content and automatically returns the name, severity, owner, and description — without any Parse JSON action.
LLM models available in inline agents
When configuring a custom inline agent, multiple models are selectable. The exact availability depends on your environment, but the range covered at the time of writing goes from Claude Sonnet to models from the GPT-5 family. For classification or structured extraction tasks, GPT-5 chat is a balanced choice between performance and cost.
Tools available in an inline agent
An inline agent can connect to the following services via the Tools tab: Outlook, Teams Messages, SharePoint. Dataverse has migrated from the Knowledge tab to the Tools tab — remember to update your existing configurations if you're migrating from the old Copilot Studio experience.
Human assistance and managing edge cases
Inline agents have a Request for human assistance option. When enabled, the agent automatically adds the request human assistance tool to its configuration. In case of a blockage or ambiguity, the agent can escalate to a human operator without uncontrollably interrupting the workflow.
This feature is particularly relevant in critical ticket processing scenarios or workflows involving high-impact decisions.
Comparison: inline agent vs published agent in a workflow
| Criterion | Inline agent | Published agent |
|---|---|---|
| Visibility in Copilot Studio | Not visible | Visible in the catalog |
| Reusable in other workflows | No | Yes |
| Typical use case | Workflow-specific logic | Cross-functional and shared skills |
| Publication cycle required | No | Yes (publication mandatory) |
| Structured output available | Yes | Yes |
Key takeaways
- Inline agents exist only within the context of the workflow that hosts them: they don't clutter the Copilot Studio catalog.
- The Show conversation in M365 Copilot option allows you to view results directly in the Copilot interface, independent of the workflow.
- Structured output advantageously replaces
Parse JSONactions for data extraction scenarios from an LLM response. - The triggers available in workflows go well beyond the manual trigger: recurrence, Teams messages, emails, SharePoint file drops — all entry points for automating repetitive tasks.
- Dataverse is now a tool, not a knowledge source: verify your configurations if you're migrating from the old Copilot Studio experience.
- Only published agents can be referenced from an external workflow.
