◈ 01 / The Stack

Seven planes, one assembly line

The atlas reads top to bottom. Two platform bands sit above the four core layers, one sits below. Each block answers a single builder question and carries its own slice of the Build 2026 announcements.

CONTROL PLANE ⛒

Productionisation & Governance

Can this agent be trusted near real users and enterprise data?

The controls you wrap around an agent before it ships: identity, session state, evaluation, observability, and ROI. This is what turns a demo into something an enterprise will actually run.

Runtime / session state
Agent Service hosted agents
Azure Functions serverless agents
Container Apps sandboxes
Agent Identity
ACS
ASSERT
Rubric
Evals / Tracing
Observability
ROI
Agent 365
Purview / Defender / Entra
Agent distribution
Windows agent containment
Discovery evidence loops

In journeys — J1 J3 J4 J5 J6

BUILD PLANE ⌗

Dev / Backend Builder Substrate

What do I actually build the sample with?

The frameworks, SDKs, local loops, and MCP surfaces that turn an announcement into working code. This plane is where this repo lives.

Agent Framework
Foundry Toolkit for VS Code
Azure Functions agents / MCP
Rayfin
HorizonDB
Cosmos DB MCP
ACA samples
GitHub SDKs / sandboxes
GitHub Copilot app
Rayfin CLI
Visual Studio AI loop
Windows AI dev loop

In journeys — J1 J2

1 MODEL

Model

What intelligence do I use?

The reasoning core, and how you source and adapt it — frontier hosted models, third-party catalogs, managed compute, fine-tuning, and local inference.

MAI models
Fireworks AI on Foundry
Managed Compute
Frontier Tuning
Foundry Local inference
Foundry Local on Azure Local
Model router
Task-specialized MAI family
NVIDIA / Anyscale placement

Sample touchpoint — Model router and inference substrate

2 CONTEXT

Context

What does the agent know?

The grounding layer — enterprise, operational, and web knowledge brought to the model through retrieval, graphs, ontologies, and memory.

Foundry IQ
Web IQ
Microsoft IQ
Scout
Work IQ
Fabric IQ
HorizonDB operational data
Rayfin app context
Graph / ontology / semantic context
Content Understanding
Service memory

Sample touchpoint — Foundry IQ grounding

In journeys — J1 J4 J6

3 TOOLS

Tools

What can the agent do?

The action surface — safe, governed ways to call functions, run code, search tools, and reach the outside world, mostly through MCP.

Foundry Toolboxes
Tool Search
Skills
MCP
Azure Functions MCP
Code Interpreter
Voice Live
Routines
Publishing routes
Rayfin backend generation
Browser automation (Playwright workspaces)
Windows MXC skills

Sample touchpoint — Agent Framework + Toolbox, Copilot workflow

In journeys — J1 J2 J4

4 HARNESS

Harness

How does it run safely in production?

The runtime wrapper that gives the agent identity, state, traceability, evaluation, and lifecycle controls — the bridge to the control plane.

Agent Framework
Agent Service
Identity
ACS
ASSERT
Rubric
Tracing / evals
Optimizer
ROI
Agent 365
Governed distribution
Windows MXC policy
Windows 365 Agents
Discovery review loop

Sample touchpoint — Serverless agent, Sandbox, Eval loop, Windows local-agent harness

In journeys — J1 J3 J5

METAL PLANE ▦

AI Infrastructure

Where does the workload actually run?

The compute and execution substrate — local, cloud, serverless, sandboxed, and accelerated silicon that make the demos runnable.

Cobalt 200 VMs
NVIDIA Vera / Rubin NVL72
RTX Spark → DGX Station for Windows
Majorana 2 quantum chip
Foundry Local
Managed Compute GPUs
Windows AI APIs
Surface RTX Spark Dev Box
Windows 365 Agents
Secure sandbox execution

In journeys — J3

Now see how the layers become builds.

Browse the sample tracks →