Microsoft’s Next Lesson: Real Open Agentic Interfaces or Bust

I love what Microsoft has done with Copilot and the Azure AI Stack—but if we’re being honest, the walled gardens are holding devs back. Meanwhile, OpenAI and Anthropic are racing ahead with open agentic tool APIs, flexible sandboxing, and composable, scriptable agent workflows. You can wire Claude Code or GPT-4o into your own tool servers, define custom evals, and hack together bespoke loops. That’s the difference between an AI assistant and a true AI co-creator.

Here’s what Microsoft should steal, stat: first, expose genuine agentic interfaces in Copilot. I want to wire my own tools (or those my team already uses) into Copilot’s action loop—not just pick from a static menu. Let us define workflows as code, not as blobs in a config GUI. Second, take Windows’ AI APIs further: add support for 3rd-party agent runners and compositional toolchains. The days of monolithic, closed ‘skills’ or actions are over. Devs want to mix, match, and own their agentic logic.

Azure is in a unique spot: the infra muscle is there, but the developer-facing APIs are still too abstracted and proprietary. Look at what OpenAI’s new Action API or Anthropic’s Tool Server SDK enable: real plug-and-play, testable, and inspectable agent workflows. Microsoft should match this openness, or risk Copilot becoming a smart but inflexible sidecar.

The bottom line: engineers need agentic systems they can shape, debug, and extend—not just consume. Microsoft’s next leap isn’t more features, it’s more control and composability. Let’s learn from the open agentic movement, before devs move on without us.