What shipped in 2026, and where it lands

This card exists because a study guide written against a fast-moving product goes
stale quietly. Microsoft shipped a great deal into Copilot Studio and Microsoft
365 Copilot during 2026. Here is what landed, what status it is at, and which
card it belongs to.

Generally available — safe exam territory

Extensibility. A2A is generally available (April) — agent-to-agent,
including across vendors. Computer Use is generally available (May), so it is
no longer a preview curiosity: agents can drive browsers and desktop applications
on a user's behalf. And the Windows 365 for Agents MCP server went GA in June,
giving an agent full operational control of a cloud PC including desktop
interaction, browser automation and semantic UI inspection. Cards 3, 19, 22.

Testing and evaluation — the biggest practical change. Agent evaluations
went GA in March with customisable test sets. Multi-turn conversation tests
followed, so you can evaluate realistic dialogue rather than single turns. And
evaluations can now be triggered from the Power Automate connector, with a
REST API in preview for CI/CD. Cards 25, 26, 41.

Models. Claude Sonnet 5 and GPT-5.5 Chat became generally available as
an agent's primary model in June, alongside options from Anthropic, xAI, Mistral
and OpenAI. One caveat that matters more than the names: some models require
data processing outside your organisation's geography. Cards 10, 33.

Operate and govern. The agent inventory schema lets you discover and audit
every Copilot Studio agent from the admin center, an API, or Azure Resource Graph.
The Analytics Viewer role grants analytics access without broader maker
permissions. Activity maps show an agent's sequence of inputs, decisions and
outputs for diagnosis. And agent flows can now respond asynchronously, past
the old two-minute limit. Cards 22, 27, 32, 42.

Still preview — know it, don't lead with it

Build. The GitHub Copilot harness arrived in June as a production-ready
preview — card 43 covers it properly. Connected agents let an agent delegate
to specialists, so you can build a single front-door agent that routes. Foundry
IQ can be used as a knowledge source directly from Copilot Studio.

Govern. Entra Agent IDs can be created automatically for each agent, which
scopes connector permissions, Conditional Access policies and DLP governance to an
individual agent rather than to the maker. That is a meaningful tightening of card
42's story. An agent readiness status page consolidates publishing errors,
runtime issues and configuration blocks with severity levels.

Voice and channels. Teams Phone Agent can hand a call off to your custom
voice agent. Consent-based recording asks callers for consent with
configurable compliance behaviour and retention. Hold and resume lets a voice
agent pause mid-conversation. Card 40.

Microsoft 365 Copilot side. The Agent Registration API and Package
Management API let administrators register, block, unblock and reassign agents
programmatically. The Copilot policy settings API gives one endpoint across
Cloud Policy Service and Intune. Cards 20, 21.

The three that change an answer, not just your vocabulary

Agent Builder is narrowing, not widening. Learn still says plainly: if you
need Actions to integrate external services, use Copilot Studio. But the
declarative agent manifest gained EmailActions (triage, supervised send,
delete, inbox rules, auto-reply, folder management) and MeetingActions
(scheduling, time-finding polls, time insights) in version 1.8. So the rule "Agent
Builder advises, it never acts" now holds specifically for external services —
first-party Microsoft 365 actions are the exception. Worth knowing precisely,
because it is the sort of distinction a well-written question turns on.

Evaluation has become a pipeline step. GA evaluations, plus the Power Automate
connector, plus the REST API, mean an answer that gates a release on an evaluation
run is now something you can actually build. It belongs in your ALM design, not in
a testing afterthought.

Grounding got a new front door. Bing Custom Search can be added as a
knowledge source using a Custom Configuration ID — a curated, scoped web index
rather than the open web. Narrower and considerably more defensible when a
scenario worries about what the agent might read.

Exam trap. The exam page says most questions cover generally available
features, with preview only where commonly used. Use the first section with
confidence and treat the second as product literacy rather than revision.

The one line to carry: naming a specific model is rarely the answer — the
reasoning about cost, latency, residency and quality is, because the names on that
list changed four times during 2026 alone.

Goes with: card 43 for the harness restructure, and REFRESH.md in the study
repo for how this card gets kept honest.

---

Updated 23 August 2026

The weekly Microsoft Learn check moved two pages. One of them changed an answer.

Entra Agent ID left preview and became mandatory. From July 2026 Copilot Studio
creates one automatically for every new agent and the environment-level opt-out is
gone. Earlier versions of this card listed it under preview; it now belongs in the
generally-available column, and "turn on agent identities" has stopped being a valid
recommendation. Card 42 has the detail.

Three additions on the GitHub Copilot harness, all preview: a workflow or an
MCP server can be added as a tool, users can attach files to a conversation and
see the files the agent creates, and MCP servers can be submitted for Microsoft
certification — a published signal that a server meets Microsoft's expectations for
reliability, security, compliance and responsible operation. That last one is a trust
marker for third-party tooling rather than a technical capability, which is exactly
how a question would test whether you understand it.

And one non-change worth stating. The Agent 365 page was revised too, but nothing
this card asserts about it moved — generally available since 1 May 2026, per user,
works best on E5, registry and Agent Map. A date changing is not the same as a fact
changing, and saying so is part of keeping the card honest.