Part 1 · Why was it rebuilt?
Classic's structural limits and why the core was rebuilt.
0. Copilot Studio: why rebuild it?
▶ Point — Classic was a tool for “designing conversations,” and that paradigm itself hit a ceiling for multi-step, long-running work. So Microsoft rebuilt not the UI, but the AI core itself on top of a coding harness.
0.1 Classic’s structural limits
Classic Copilot Studio (formerly Power Virtual Agents) was essentially a conversation flow design tool. Makers had to draw these elements directly:
- Topics — trigger-response trees such as “if the user says X → node A → node B”
- Prompts — one-off AI calls from a specific input to a specific output
- Branches, conditions, and variables — connecting every possible case as nodes
The decisive weakness of this approach is that people must imagine and draw every path in advance. As soon as a scenario becomes even slightly complex, branches explode; when users say something unexpected, the flow breaks.
This model fit FAQ chatbots and structured scenarios well, but its limits were clear.
| Limitation | Symptom |
|---|---|
| Flow breaks | An unexpected input in the middle of a multi-step task cuts off the topic |
| Branch explosion | As scenarios become more complex, nodes and conditions increase exponentially |
| Rigid tool calls | The maker must predesign which tool to call and when |
| Long work is difficult | Long-horizon tasks such as “read these documents, analyze them, and write a report” are hard |
Anthropic summarized exactly how these limits surface in complex work as two failure modes in Effective harnesses for long-running agents—the same patterns that made Classic bots fall apart.
| Failure mode diagnosed by Anthropic | Symptom in Classic bots |
|---|---|
| Tries to do everything in one shot and runs out of context halfway through | A multi-step task stops midway and ends incomplete |
| Reports progress and declares the job done too early | The conversation ends even though core functionality does not work |
0.2 What changed — rebuilding the core
Customer feedback centered on three points (based on slide 4 of the blog and roadmap materials):
- They want agents to become a natural part of work
- They need a reliable and scalable agent platform
- They want agents, apps, and deterministic workflows to work well together
So Microsoft’s choice was not to tweak the UI, but to rebuild the AI core (orchestrator) on a new coding harness and CLI layer. The resulting capabilities are:
- Stronger instruction adherence — especially retaining the original instructions even in long conversations
- Long-horizon, multi-step work — continuing to completion without collapsing midway
- Recursive task execution — breaking complex problems into steps on its own and iterating
- Large content + rich file generation — producing Word, PowerPoint, Excel, CSV, and PDF outputs in a dedicated container
The build experience was also simplified: configuration tabs went from nine to four (Build, Test, Preview, Monitor); instructions, skills, tools, and knowledge are on one screen; and in full-page testing, you can observe the reasoning process (inline chain-of-thought) and tool calling in real time.
FAQ response: “Isn’t this just a UI reskin?” → The improvements in reliability and multi-step processing come from the new AI core, not the UI. The redesigned UI simply improves the maker authoring experience that supports that core.