Skip to content
agents.cli

Course · Copilot · The modes

The modes

Every Copilot chat can run at more than one level of latitude. The same assistant, reading the same project, can be answering questions without touching a file, laying out an approach for your approval, or editing across the repository and running commands on its own. Which level you are in is set by the agents dropdown in the Chat view (if you have not opened the Chat view yet, Setup covers it), and the built-in entries are three: Ask, Plan, and Agent - the “chat modes” of Copilot’s documentation; this course says role for one entry and modes for the set.

Choosing between them is not a preference setting. It is the core decision of the control principle this course frames everything with - control being what guides execution and enforces boundaries - and it comes down to two questions: how much of the work do you hand over, and what do you keep? This chapter is where that judgment gets built, because every later chapter - rules, permissions, delegation to the cloud - assumes you can already make it.

The three built-in roles

Section titled “The three built-in roles”

Each role is the same Copilot reading the same project. What changes is how far it goes on its own:

  • Ask - reads and explains, changes nothing. The way to understand unfamiliar code before touching it.
  • Plan - researches and produces a step-by-step approach, then stops for your review before code exists.
  • Agent - runs the whole loop: plans, edits across the files it decides are relevant, runs commands and tests, and self-corrects until it judges the task done.

Those are the behavior guarantees this chapter teaches by, as documented in September 2026; the Copilot documentation overview tracks the current set.

If you learned a fourth mode, Edit, from an older guide: it is gone from the dropdown, and Scoped edits in Agent mode covers what replaced it.

The natural rhythm on real work runs through the three in order: ask to understand, plan to agree the approach, let Agent build. Switching mid-task keeps the conversation - your questions, its findings - carried forward into the next role, so climbing from one role to the next is not three separate chats. It is one piece of work, widening the agent’s latitude as your confidence grows.

The lessons below meet Ask first, then the two ways to run Agent - scoped, then the full handoff - and only then Plan, because Plan’s approve-before-build gate is easiest to understand once you have seen the autonomous run it gates. On a real task, the sequence above is the one you run.

The running example

Section titled “The running example”

The lessons ahead use one small scenario so the choices have stakes. You have inherited two TypeScript repositories: orders-service, an order-management API your squad owns, and shared-lib, an internal library it imports along with roughly a dozen other services. A feature lands on your desk: orders above a configurable threshold must require a second approver. As you explore, it turns out the threshold check and the audit helper it needs live in shared-lib - so one feature becomes two changes with very different risk: a careful library change, and a routine change in your own service.

The scenario is deliberately small; substitute any service you own plus any library other teams consume. Each page restates what it needs. What the example supplies is not the code but the two risk profiles, because the role you reach for is decided by risk, and this chapter puts the same task in front of both profiles.

The lessons

Section titled “The lessons”

By the end, the goal is not knowing what the three roles do. Anyone can read three definitions. It is glancing at a task and knowing how much latitude it is asking for.