microsoft​-build-2026 REFRESHED 2026.06.06

03 / Builder Journeys

Six builds, read left to right

The stack reads top to bottom — layers you assemble. Journeys read the other way: six end-to-end workflows that cut across the stack, each composed from existing sample tracks and demo narratives. The question they answer is simply — what would I actually build, end to end?

Narrative Sketch Runnable later
J1

Prompt-to-Production Agent

Narrative

What you build A local agent prototype that ends up grounded, governed, evaluated, and published — one continuous path from prompt to production.

Agent Framework Foundry IQ Toolboxes / MCP Agent Service Evals Publish
Layers touched
  • Build
  • Context
  • Tools
  • Harness
  • Control

What you learn How grounding, governed tools, and evals compose into one shippable agent instead of four disconnected demos.

When you'd reach for this You have a working prototype and need the path that turns it into something an enterprise will actually run.

J2

Agentic Developer Workflow

Sketch

What you build An issue that becomes a reviewed, deployable change — the coding agent drives the loop while you keep the gates.

GitHub Copilot app Worktrees Rayfin backend PR review Deploy
Layers touched
  • Build
  • Tools
Composed from track-10 ↗ track-08 ↗

What you learn Where the human checkpoints sit in an agent-driven dev loop, and how generated backends stay traceable.

When you'd reach for this You want coding agents to do more of the build without losing review, rollback, or attribution.

J3

Secure Local Agent

Narrative

What you build An agent that runs on local / edge hardware under containment, identity, and policy — then graduates to a managed Cloud PC.

Windows AI / Foundry Local MXC containment Identity attribution Agent 365 Windows 365 for Agents
Layers touched
  • Harness
  • Metal
  • Control
Composed from track-09 ↗ track-07 ↗

What you learn How local execution gets bounded — sandboxing, attribution, and audit — before it ever touches enterprise data.

When you'd reach for this You need an agent close to local files or devices, but with enterprise containment and a path off the laptop.

J4

Governed Enterprise Action Plane

Narrative

What you build Work knowledge plus governed connectors and a managed gateway that turn an agent into something allowed to take action.

Work IQ Logic Apps Connector Namespace / hosted MCP APIM / API Center Audit
Layers touched
  • Context
  • Tools
  • Control
Composed from track-03 ↗ track-01 ↗

What you learn How actions get governed at the plane level — connectors, hosted MCP, and an API gateway in front of every tool.

When you'd reach for this Your agent has to do real work across enterprise systems, and every action needs a policy and an audit trail.

J5

Safe Code Execution + Evaluation

Sketch

What you build Agent-written code that runs in a disposable sandbox, then gets scored and watched — execution and evaluation as one loop.

Container Apps session sandbox ASSERT ACS Rubric Observability
Layers touched
  • Harness
  • Control
Composed from track-04 ↗ track-05 ↗

What you learn How sandboxed execution, policy specs, and rubric evals combine into provable safety, not a demo that worked once.

When you'd reach for this You're letting an agent run code and need to show it was contained, scored, and observable.

J6

Research / Evidence Agent

Narrative

What you build A research agent that proposes, gathers evidence, and logs every claim — with a human reviewer in the loop before handoff.

Microsoft Discovery Hypothesis Evidence loop Source registry Reviewer checkpoint
Layers touched
  • Control
  • Context
Composed from track-11 ↗

What you learn How an evidence-first agent keeps citations, uncertainty, and reviewer checkpoints auditable end to end.

When you'd reach for this You want an agent for R&D or analysis whose conclusions you can trace back to sources and sign off on.