Insight · Safe AI
The new Copilot Studio: skills, harnesses and the controls that still matter.
Microsoft rebuilt Copilot Studio this year. Agents on the new GitHub Copilot harness have no topics, trigger phrases or node trees. You describe the agent in plain language, attach skills and tools, and an orchestrator decides what to run and in which order.
Makers get speed. The people who approve, audit and pay for agents get new control points. Four matter most.
The skill description now decides what runs. A skill is a reusable package of written instructions: prerequisites, steps, validation rules and the response format. The orchestrator reads each skill’s short description and matches it against the request. A vague description sends a request into the wrong business process, so review descriptions with the care you give a flow’s trigger condition.
Skills are shared assets. One skill can be versioned, reused across several agents and imported from other tools. Give each skill an owner, a change process and an approval record, as you would a shared connector.
Instructions guide. Controls enforce. A line such as “never display employee ID numbers” steers the model but enforces nothing. Data loss prevention policies, connector restrictions, environment security and permissions on the underlying data do the enforcing. Write the instruction, then put the real control underneath it.
Memory and cost are new line items. Agents on the new harness can keep memory across sessions. That memory is stored information, so it belongs in your retention and privacy review under PIPEDA and Quebec Law 25. Billing changed too. GitHub Copilot harness agents consume Copilot Credits from the first build session, including previews, tests and evaluations. An unmonitored maker environment can run up a bill before anything goes live.
Decide the harness before the build. Agents cannot move between the new harness and the standard one. If a standard harness agent already answers questions predictably, rebuilding it may cost more than it returns.
The four accountability questions still apply: who can access the agent and its data, what it can do, who approved it, and who answers when it acts. The foundation for answering them is the same one covered in governing AI agents before they act.
For the technical detail, two articles by Muhammad Ahmed on LinkedIn: Where did Topics go? covers skill authoring with a complete example, and Demystifying the GitHub Copilot harness compares the three harnesses.