Agent behaviours, and agents inside Microsoft 365

Two subjects on one card: behaviours you switch on deliberately, and what changes
when an agent lives inside Microsoft 365.

Behaviours you design deliberately

Reasoning. The agent plans multi-step work instead of answering in one shot.
Better on complex asks, and slower and dearer. Turn it on for genuinely
multi-step scenarios rather than across the whole agent.

Voice mode. Spoken interaction is a different design, not the same design with
a different skin. Shorter turns, no bullet lists, and it has to handle being
interrupted. Confirmations matter far more when nothing is on screen to check
against.

Knowing when to stop. Explicit fallback and escalation. Say it does not know
rather than improvise. And the line worth remembering because it cuts the other
way: over-refusal is also a failure. An agent that refuses everything is safe
and useless, and the exam does treat that as a wrong answer.

Agents in Teams and SharePoint

Teams. Meets people where they already talk. Good for knowledge and light task
agents. Group context means being careful about what gets disclosed into a
channel where everyone can see it.

SharePoint. Agents grounded on a site's content, where permissions are
inherited, not invented. Which means existing oversharing becomes instantly
discoverable — the same point as card 5, arriving from a different direction.

Optimising the design. Put the agent where the work happens, not where it is
easiest to build. One surface per audience, rather than forcing everyone into a
single channel because it was convenient to deploy there.

Exam trap. Turning on reasoning or voice looks like an upgrade and costs like
one. Reasoning multiplies tokens and latency on every turn — the right answer
enables it for the subset of scenarios that need multi-step planning, not across
the whole agent.

The one line to carry: reasoning and voice are not upgrades, they are trades —
switch them on where the scenario needs them and nowhere else.

Goes with: card 40, which covers the two kinds of voice agent and what
reasoning actually costs when you switch it on.

---

Currency (19 Aug 2026). On the Microsoft 365 side: the Agent Registration API
and Package Management API let administrators register, block, unblock and reassign
agents programmatically, and the Copilot policy settings API gives one endpoint
across Cloud Policy Service and Intune. All preview. See card 44.