On October 21, Microsoft posted MC1518233, Copilot Studio: Enhanced support for Copilot Chat agents to the Message Center. Copilot Chat agents, also called declarative agents, are the lightweight agents people reach directly inside Microsoft 365 Copilot Chat. With this update they gain six capabilities that until now belonged to full Copilot Studio agents: topics, connected agents, evaluations, monitoring, a native Copilot Chat preview with an Activity Map, and lifecycle management across development, test, and production.
The post is marked for awareness, with no action required, and general availability was scheduled for October 26, 2026. That label is accurate for the tenant setting and misleading for the governance model. Many organizations govern declarative agents lightly because they have been the simple tier. After this update, a Copilot Chat agent can run deterministic flows, delegate work to other agents, and move through environments like any production solution. The tier assumption is worth revisiting.
The update at a glance
| Message Center ID | MC1518233 |
| Published | October 21, 2026 |
| Rollout | General availability, October 26, 2026 |
| Applies to | Copilot Chat agents (declarative agents) built in Copilot Studio |
| Admin action | None required; Microsoft recommends updating internal guidance and telling makers |
Where Copilot Chat agents sit today
Microsoft offers three ways to build a declarative agent, and its tool comparison page describes them side by side. The differences matter for governance because they determine who can build, what the agent can reach, and how changes are controlled.
| Agent Builder | Copilot Studio | Microsoft 365 Agents Toolkit | |
|---|---|---|---|
| Approach | No code, natural language | Low code | Pro code in VS Code, Visual Studio, or the CLI |
| Maker license | Microsoft 365 subscription | Microsoft 365 subscription plus a Copilot Studio license | Microsoft 365 subscription with sideloading enabled |
| Actions | Not for complex workflows | Power Platform connectors and custom API actions | Custom API actions from OpenAPI specs; no Power Platform connectors |
| Lifecycle | Submit to the organization’s catalog | Power Platform pipelines, native Git integration, Copilot ALM Starter template | Source control, pull requests, and CI/CD |
MC1518233 widens the gap between the Agent Builder and Copilot Studio columns. A Copilot Chat agent built in Copilot Studio now carries most of the machinery of a full agent while still appearing to users as a simple agent in Copilot Chat. Governance should follow what the agent can do, not which builder produced it.
Six capabilities, six governance decisions
1. Topics for deterministic conversational flows
Topics let a maker script exactly how the agent responds to a defined set of questions instead of leaving every answer to generative orchestration. Microsoft’s Message Center examples are standardized HR answers and content that has to follow brand and legal requirements. That is the strongest part of this update: for questions where the wording is policy, a scripted answer is more defensible than a generated one.
The decision is ownership. A topic that answers a leave policy question is a published statement of HR policy, so it needs a named content owner, a review cycle, and a way to retire it when the policy changes. Treat topics in Copilot Chat agents as controlled content, not as agent configuration.
2. Connected agents
Connected agents let one Copilot Chat agent hand part of a task to another. Microsoft’s documentation, Connect to other agents from a declarative agent, sets out the rules. The connected agent can be another declarative agent or an agent with a registered Agent2Agent (A2A) card. It must be registered in the Microsoft Agent 365 agent registry, and users must have access to it, either by acquiring it or through admin deployment. Copilot chooses which connected agent to call based on its name, description, and example prompts, and the exchange between agents is text.
One detail deserves attention. When a connected agent calls an API plugin that requires user confirmation, the documentation says any Adaptive Cards go to the active agent rather than to the user. The user does not see the card or a dialog; the active agent processes the data. Review any connected agent that relies on confirmation prompts as a control before you allow it to be called from another agent. Also note that descriptions now act as routing logic, so a vague description is a quality and safety problem.
3. Evaluations
Makers can now run evaluations against Copilot Chat agents. Combined with the Review panel and the expanded evaluation metrics covered in PPCC 2026: Every Announcement, and the Governance Decision Behind Each One, this gives organizations a consistent way to test the agents most employees will actually use.
The decision is whether evaluation is optional or required. For any agent that answers policy, HR, legal, or finance questions, require a test set that covers the scripted topics and the most likely questions outside them, and rerun it when the underlying model or knowledge changes.
4. Monitoring
Monitoring gives makers visibility into how a Copilot Chat agent is used. The Message Center post does not list which metrics are included, so confirm what your tenant shows before you build reporting around it. What changes is that usage of declarative agents becomes something you can measure, which makes retirement decisions possible. An agent nobody uses is still an agent someone has to maintain, and a regular review of low-usage agents is one of the simplest ways to limit agent sprawl.
5. Native Copilot Chat preview with Activity Map
Makers can test the agent as users will see it in Copilot Chat, with an Activity Map showing each step. Microsoft’s activity map documentation describes how the feature works for Copilot Studio agents generally: each node records inputs, decisions, and outputs, such as the search query and sources behind a knowledge answer. The same page sets out governance-relevant defaults. Makers see only their own interactions unless an admin enables sharing of activity transcripts. Historical activity requires an Exchange license and mailbox, and it is stored through Microsoft 365 services in the location associated with that mailbox.
Decide who should be able to review other people’s conversations with an agent before anyone asks to turn transcript sharing on. That is a privacy decision as much as a debugging one.
6. Lifecycle management across development, test, and production
Copilot Chat agents can now be deployed as separate versions across development, test, and production. For Copilot Studio, Microsoft’s tool comparison lists Power Platform pipelines, native Git integration, and the Copilot ALM Starter template as the lifecycle tools.
This is the capability that most changes the governance picture, because it brings declarative agents into your environment strategy. If your organization builds Copilot Chat agents in the default environment today, this is the moment to stop. Copilot Studio Mistake: No Environment and ALM Strategy covers the environment pattern, and The One Switch That Decides Whether Your Pipelines Are Governed covers the pipeline setting that keeps promotion under admin control.
What this means for cost
Microsoft’s Copilot Studio billing page sets out the rule that matters most here. Employee-facing use of Copilot Studio agents is included in the Microsoft 365 Copilot user license when the person using the agent holds that license and the agent runs under their authenticated identity, subject to fair usage limits. Users without that license consume Copilot Credits at the standard rates.
| Feature | Unlicensed user | Microsoft 365 Copilot licensed user |
|---|---|---|
| Classic answer, including predefined topic responses | 1 Copilot Credit | No charge |
| Generative answer | 2 Copilot Credits | No charge |
| Tenant graph grounding | 10 Copilot Credits | No charge |
| Agent action, including topic transitions | 5 Copilot Credits | No charge |
| Agent flow actions, per 100 actions | 13 Copilot Credits | No charge when called through the agent trigger |
Two consequences follow. First, a Copilot Chat agent rolled out to staff without Microsoft 365 Copilot licenses is a metered agent, so it needs a cost owner and an environment with credit allocation, as covered in From Copilot Credits to CFO Dollars. Second, the billing page lists topic responses as classic answers, so moving frequently asked policy questions into topics may lower the per-response rate for unlicensed users as well as improving consistency. The billing page does not address declarative agents in Copilot Chat by name, so confirm this in your own usage reports before you build a business case on it.
A checklist for admins
- Inventory the Copilot Chat agents already in your tenant, and note which were built in Agent Builder and which in Copilot Studio.
- Classify agents by what they do, not by builder. Any agent that answers policy questions or calls connected agents belongs in your governed tier.
- Assign a content owner to every topic that states HR, legal, brand, or finance policy, with a review date.
- Decide who may register agents in the Agent 365 registry for use as connected agents, and review any that depend on confirmation prompts.
- Require an evaluation test set before a governed Copilot Chat agent is published, and rerun it after model or knowledge changes.
- Decide whether activity transcript sharing should be enabled, and for whom.
- Move Copilot Chat agent development out of the default environment and into a development, test, and production path with pipelines.
- Identify which audiences lack Microsoft 365 Copilot licenses, and give every agent they use a cost owner.
- Update maker guidance and training, which is the one recommendation Microsoft makes in the Message Center post.
What is still unclear
As of November 2, 2026, Microsoft’s public documentation has not caught up with the Message Center post. The Copilot Studio what’s new page has no October entry for this change. The connected agents documentation does not state a preview or general availability status. The activity map page lists the Microsoft Teams and Microsoft Copilot channels but does not say whether every detail applies to declarative agents. The post also does not say which monitoring metrics are included or how version promotion works for these agents. Expect the Learn documentation to fill these gaps, and check it before you write internal policy that depends on the details.
The direction is clear even where the details are not. The line between a simple Copilot Chat agent and a production Copilot Studio agent is fading, and the governance model that treated them differently should fade with it.
Sources
- MC1518233: Copilot Studio, Enhanced support for Copilot Chat agents, Microsoft 365 Message Center, October 21, 2026
- Connect to other agents from a declarative agent, Microsoft Learn
- Declarative agent tool comparison, Microsoft Learn
- Copilot Studio activity map and agent activity, Microsoft Learn
- Copilot Studio billing rates and Copilot Credits, Microsoft Learn
- What’s new in Copilot Studio, Microsoft Learn