ai-integration-bundle

# AI Integration Bundle Agent uses this bundle for AI requests that span Windows on-device inference, Foundry Local, MCP-connected tools, or Azure AI cloud services. ## When to Use | Prompt Pattern | Use This Bundle | |---|---| | Runtime selection across device, local model catalog, and cloud | Yes | | AI workflow needs MCP-aware tool orchestration | Yes | | Local-first plus cloud fallback design | Yes | | One AI surface is already fixed and no routing choice remains | No | ## When Not to Use | Prompt Pattern | Agent Route | |---|---| | Direct Windows inference API work only | `windows-ai-apis` | | Foundry Local, Windows connector, or ODR discovery flow only | `windows-ai-mcp-integration` | | Azure-hosted model or managed AI service only | `azure-ai-services` | ## Required Inputs | Input | Required | Notes | |---|---|---| | Inference location | Yes | On-device, cloud, or hybrid | | Connectivity expectation | No | Offline, connected, or intermittent | | Data sensitivity | No | Local-only, regulated, or cloud-approved | | Tooling pattern | No | App API, MCP server, managed endpoint, or hybrid | ## Activation | Request Signal | Use Bundle | Route Detail | |---|---|---| | Need WinML, DirectML, ONNX Runtime, or local model sessions | Yes | Agent routes to `windows-ai-apis` | | Need Foundry Local model catalog, prompt loop, or local LLM setup | Yes | Agent routes to `windows-ai-mcp-integration` | | Need MCP tool exposure, registry discovery, or connector-mediated file access | Yes | Agent routes to `windows-ai-mcp-integration` | | Need Azure OpenAI, Search, Speech, Vision, or managed safety controls | Yes | Agent routes to `azure-ai-services` | | Need offline-first local inference with cloud fallback | Yes | Agent combines `windows-ai-apis` or `windows-ai-mcp-integration` with `azure-ai-services` | ## Coverage Matrix | Need | Primary Skill | Primary Outcome | |---|---|---| | Local inference runtime selection | `windows-ai-apis` | Explicit WinML, DirectML, or ONNX execution path | | Foundry Local setup and local model catalog use | `windows-ai-mcp-integration` | Alias-based local model workflow with on-device execution | | MCP connector, registry, and tool-contract wiring | `windows-ai-mcp-integration` | Stable tool surface across host and local AI runtime | | Managed cloud AI services | `azure-ai-services` | Azure-hosted model, search, speech, vision, or document pipeline | | Hybrid AI architecture | This bundle | Local and cloud tasks split by latency, privacy, and scale | ## Decision Order | Decision | Agent Action | |---|---| | Offline execution, local privacy, or device latency dominates | Agent selects `windows-ai-apis` or `windows-ai-mcp-integration` first. | | Foundry Local model lifecycle or connector discovery dominates | Agent selects `windows-ai-mcp-integration` first. | | Enterprise scale, hosted models, or managed safety dominates | Agent selects `azure-ai-services` first. | | Tool interoperability matters | Agent adds MCP after the runtime choice is explicit. | | Hybrid execution remains in scope | Agent assigns each task to one local or cloud owner before implementation. | ## Workflow | Step | Agent Action | Test | Pass | |---|---|---|---| | 1. Classify runtime | Agent maps the request to local, cloud, MCP, or hybrid tracks. | Agent lists one or more tracks. | Every requested capability maps to one track. | | 2. Select the primary runtime | Agent chooses Windows, Foundry Local, or Azure as the first implementation surface. | Agent records one primary runtime. | Runtime choice fits latency, privacy, and connectivity inputs. | | 3. Add MCP only when needed | Agent maps tool hosts, connector boundaries, and schema ownership. | Agent records provider and consumer roles. | MCP appears only where tool interoperability adds value. | | 4. Route to specialist skills | Agent reads only the routed skills and writes the smallest aligned plan or change. | Agent reviews scope. | Output stays inside the routed surfaces. | | 5. Verify operational fit | Agent checks privacy, latency, scale, and fallback behavior. | Agent compares architecture to stated inputs. | Selected pattern fits the request constraints. | ## Verification Matrix | Track | Agent Verifies | Test | Pass | |---|---|---|---| | Windows runtime | Local inference path fits device and privacy goals. | Inspect runtime location and provider choice. | On-device execution stays intentional. | | Foundry Local | Local model alias and prompt loop fit the task. | Inspect alias, catalog, and local prompt path. | Model execution stays local and reproducible. | | Azure AI | Cloud path fits managed scale and service operations. | Inspect endpoint, auth, and service selection. | Azure route uses managed services intentionally. | | MCP | Tool contract is stable and runtime-aware. | Inspect tool schema and host boundary. | Tool inputs and outputs stay explicit. | | Hybrid | Local and cloud tasks have explicit boundaries. | Inspect task routing and data flow. | No task crosses boundaries without a reason. | ## Common Pitfalls | Pitfall | Agent Fix | |---|---| | Sending local-only tasks to cloud by default | Agent keeps those tasks on device. | | Treating Foundry Local as a cloud substitute with no connector or runtime plan | Agent records alias, hardware path, and invocation flow explicitly. | | Adding MCP where no tool contract exists | Agent removes MCP and keeps direct API integration. | | Using local inference for workloads that need hosted scale or managed models | Agent routes those workloads to Azure AI. | | Mixing local and cloud data paths with no ownership boundary | Agent assigns each workflow step to one runtime. |