Agent Framework · Shown at Microsoft Build
Personal agents & suggested actions
Copilot suggests an action on the data a maker is already looking at — pick recipients, set a cadence, preview the email, and a personal agent exists.
Overview
Building a Copilot agent from scratch is a lot to ask of someone in the middle of a task. This feature flips the starting point: Copilot looks at the data a maker is already working in and suggests a specific, useful action — pick recipients, set a cadence, preview the email — so a personal agent can exist without ever opening a blank canvas.
It was a 0 → 1 concept, designed and shown publicly at Microsoft Build as part of the Agent Framework.
This sat within the broader product direction for Suggested Actions: evolving model-driven applications from data-oriented experiences toward task-oriented ones, using AI to recommend relevant actions while keeping people in control of the decision. That created two linked design problems — helping makers define useful actions without needing to understand the technical complexity behind AI-generated recommendations, and helping end users trust a recommendation without it feeling automatic or unexplained.
Design decisions
-
1
Anchored the suggestion in context — the action Copilot proposes is generated from the data already on screen, so the maker never has to describe their intent from zero.
-
2
Kept the path from suggestion to working agent to three visible steps — pick recipients, set a cadence, preview the email — so the commitment stays small at every point.
-
3
Made the preview step non-negotiable before an agent goes live, so a maker never publishes something automated without seeing exactly what it will send.
-
4
Moved the "New action" control closer to the existing action list while keeping Save tied to the full configuration — separating two different mental models: creating an individual action, and saving the overall setup.
The goal was never "AI decides what happens" — it was "AI helps a person understand what might deserve attention next," so every recommendation carried enough context for someone to evaluate it rather than blindly follow it. On the end-user side, that turned into a six-step loop:
1
Open
The maker opens the record they're already working in.
2
Evaluate
AI checks configured actions against the data on screen.
3
Recommend
A suggestion appears with context on why it may be useful.
4
Review
The person stays responsible for deciding whether it's right.
5
Act
Complete or dismiss the recommendation.
6
Feedback
That choice feeds back in to improve future suggestions.
Designing AI experiences isn't only about designing the AI response. It's about designing the relationship between human intent, system context, AI reasoning, user control, and feedback.
From configuration to a live suggestion
The maker-facing side of the same feature — turning on an app assistant agent, telling it what to suggest, and watching it show up as a Copilot Action on a real record.
1. Turn on the app assistant agent for an app.
2. Define what to suggest, and when — in plain language, not rules syntax.
3. Save, with a clear confirmation that it took effect.
4. Live — Copilot Actions now appear on the maker's own data.
A personal agent exists without a maker ever opening a blank canvas.