Copilot Chat Agents in Copilot Studio Reach GA: Topics, Connected Agents and a Real Lifecycle

If you build agents for Microsoft 365 Copilot Chat, October 26, 2026 is the date to remember. That is when Microsoft says the enhanced Copilot Studio support for Copilot Chat agents reaches general availability, according to Message Center notice MC1518233, published on October 21.

The notice is short, and the interesting part is a list of things makers could not do before in this kind of agent: deterministic flows, connected agents, evaluations, monitoring and a proper dev, test and production path. This post goes through what each one changes in practice, and where I would be careful before moving a real workload over.

What the notice says GA includes

From the Message Center post, Copilot Studio makers building Copilot Chat agents get:

• Topics for deterministic flows, so some paths are scripted instead of left to the model
• Connected agents, so one agent can hand work to another
• Evaluations, to test an agent against a set of expected behaviours
• Monitoring of what the agent actually did
• A native Copilot Chat preview with an Activity Map, to see how a conversation moved through the agent
• Lifecycle management across development, test and production environments

Microsoft marks the message as awareness only, with no admin action required. I read that as: nothing is switched on for you, the capabilities simply appear in Copilot Studio for this agent type.

Topics: the part that makes agents predictable

A fully generative agent decides everything at run time. That is flexible, and it is also why it is hard to promise a regulated answer. Topics give you a deterministic route for the cases where wording and order matter: a refund policy, an escalation rule, a mandatory disclaimer.

One third-party guide I read says that when a user's intent matches both a topic and a connected agent, the routing engine prefers the topic. I could not find that stated in Microsoft's own documentation, so treat it as something to test in your own agent rather than a rule to design around.

My working rule: write a topic for anything you would be unhappy to see phrased differently on two consecutive days, and leave the rest to generation.

Connected agents: peers, not children

The same guide makes a useful distinction. Child agents depend on their parent, while connected agents operate as independent peers. That matters for ownership. A connected agent can have its own owner, its own evaluations and its own release cadence.

It also matters for cost and for failure. Each hop is another model call that you pay for, and another place an answer can go wrong. If you are new to the billing model, my earlier post on Copilot Studio credits covers how usage is metered.

Evaluations and the lifecycle

The two items that change team habits are evaluations and environment promotion.

Evaluations let you keep a set of test questions with expected outcomes and run it whenever the agent changes. Combined with dev, test and production environments, you finally get what software teams take for granted: change something in dev, run the tests, promote only if they pass. I wrote about wiring that into a release process in Gate Copilot Studio Releases with the Evaluation API.

Monitoring and the Activity Map close the loop. When an answer is wrong you can see which topic, tool or connected agent produced it, instead of guessing from the final text.

A short checklist before you migrate an existing agent

1. List the answers that must not vary and turn each into a topic.
2. Write ten to twenty evaluation cases from real questions, including the awkward ones, before you change anything.
3. Decide which environment is the source of truth and stop editing production directly.
4. Turn on monitoring first and read a week of conversations. You will find topics you did not know you needed.
5. Connect agents one at a time, and run the evaluations after each connection.

What I could not verify

The notice describes capabilities, not limits. I did not find published numbers for how many connected agents an agent can have, or how the Activity Map handles very long conversations. Check Microsoft Learn for the current limits before you design around a specific depth.

Sources

• MC1518233: Copilot Studio, enhanced support for Copilot Chat agents (published October 21, 2026; GA October 26, 2026)
• Copilot Studio: enhanced support for Copilot Chat agents, summary
• Multi-agent orchestration in Copilot Studio (third-party guide)