Yumin in the Loop

    All Articles

    Explore insights on generative AI, agent frameworks, and Microsoft Copilot Studio.

    An abstract document passes through a bounded decision component into three possible branches, with one branch highlighted.
    AI Trends

    Decision models: the small judgments inside useful AI workflows

    Why decision models converged so quickly across Jev, OpenAI, Perplexity, and Cloudflare, and where bounded judgments help real workflows.

    10 min read
    Nested architectural boundaries around a conversation, control loop, workspace and blue organizational identity node
    Agent Frameworks

    The Road to Autopilots Is a Stack, Not a Ladder

    A four-layer architecture model that separates chat, API agents, harness runtimes, and Microsoft Autopilots by who owns each responsibility.

    10 min read
    A white agent sandbox enclosure beside three small model modules, with one blue module fitted inside the sandbox
    Agent Frameworks

    Running local models inside Microsoft's agent sandboxes

    A small ONNX model ran locally in two Microsoft agent runtimes. The ceiling on what you can put there is packaging, not compute.

    10 min read
    Four open white runtime enclosures in different sizes, each containing blue and black compute and storage modules
    Agent Frameworks

    Four agent runtimes, measured from the inside

    Claude Cowork, Microsoft 365 Copilot Cowork, the GitHub Copilot harness, and Codex run tasks on very different machines. Here is what each runtime actually provides.

    9 min read