Copilot Studio now runs three harnesses. The GitHub Copilot harness, generally available since 3 August 2026, replaces topic-based authoring with instructions and goal-driven orchestration, and replaces per-message licensing with Copilot Credits consumed from the moment you start building. Agents cannot be transferred between harnesses in either direction. This is a rebuild decision, not an upgrade.

This session covers the mechanics.

Component model. How behaviour is distributed across instructions, knowledge, tools, memory, skills, and connected agents, and what goes wrong when the split is off. Skills are Markdown files with YAML front matter, invoked by the orchestrator on description match, portable as a SKILL.md or as a ZIP with supporting scripts and templates. We write one, then pull a second from the CAT skills gallery and adapt it.

Memory. Per-user stores in Microsoft-managed storage, invisible to the maker, deleted after 28 days of user inactivity, disabled in Teams channels and group chats, and not deleted when the toggle goes off. What that means for any agent for which continuity is a requirement, and how users inspect and clear their own memories.

Rebuild path. Live demo using Microsoft's copilot-studio-plugin with Power Platform CLI 2.9.3 to pull a standard harness agent, analyse its structure, and generate a proposed architecture on the new harness. We then review that proposal critically. Mapping every topic to a skill and every variable to memory reproduces the old design in new syntax, which is the failure mode worth naming out loud.

Evaluation and deployment. The Evaluate tab with handwritten, generated, and CSV test sets. Reading reasoning traces and error codes in Preview to diagnose failures rather than guess at them. Publishing and reading per-agent credit consumption in Monitor.

Cost. What consumes credits: LLM tokens, tools, knowledge, MCP calls, and the harness itself. Why authoring, previewing, and evaluation bill before you publish. Why a Microsoft 365 Copilot licence does not cover this harness. Capacity pack economics with no monthly rollover, the shared tenant-level credit pool, and enforcement behaviour when a tenant runs dry mid-process. Worked figures from agents I run in production.

Attendees leave with a harness selection framework, a rebuild checklist, and a cost model they can take to their own finance function.

Target audience: Copilot Studio makers, Power Platform and solution architects, AI leads, and anyone responsible for agents already running in production. Level 300.

Prerequisites: Familiarity with the classic Copilot Studio experience is helpful but not required. No coding required, though we will read Markdown skill files and run a terminal-based upgrade workflow.

Preferred duration: 45 minutes.

Demos: All live in my own tenant, against agents I run. Includes Microsoft's experimental Copilot Studio plugin for AI coding assistants (requires Power Platform CLI 2.9.3 or later). The plugin is a research project rather than a supported product, and the session treats its output accordingly. Full recorded fallback available.

Technical requirements: None if the attendees will not be building and testing

First public delivery: [fill in, or note "not yet delivered"]

Note: This platform is changing month to month. Content is verified against current Microsoft documentation ahead of every delivery, and consumption figures come from agents I run rather than from published estimates.