[Initiative] 🪼 Enable developer to build workflows

[kristenwomack]

Problem statement

azd already supports the core developer workflow from init to deploy. This pillar extends that into three areas: making azd extensible so anyone can add capabilities, supporting a complete CLI workflow for AI agent lifecycle management, and keeping developers current with safe, opt-in updates.

There are clear opportunities here. Extension developers can get stable APIs and a Go SDK instead of scripting. AI agent developers can get a seamless flow from azd ai agent init to deploy to invoke, all in one tool. And an insiders channel with azd update can give power users early access to new features with rollback protection.

This epic is focused on GA-quality for the extension framework and expanding the extensibility model framework and tooling for builders.

Vision

azd becomes the complete developer workflow surface for Azure: extensible by anyone, purpose-built for AI agent development, and always up to date. A developer can azd init, build an extension, build an agent, azd up, and be in production in minutes, as well as customize their engineering teams' custom workflows.

Who this helps

• Developers building AI agents get a seamless flow from azd ai agent init to azd ai agent deploy to azd ai agent invoke, with interactive REPL chat and streaming responses. Example: go from an idea to a running hosted agent in under 5 minutes without leaving the terminal.
• Extension developers get stable, documented APIs with a Go SDK and a clear submission pipeline. Any gRPC language works. Example: build a custom provisioning extension, test it locally, submit it, and others can azd extension install it.
• Developers using Azure AI Foundry can manage models, skills, and agents through CLI-first workflows without switching to the portal. Example: azd ai agent deploy handles container optimization, model resolution, and deployment in one command.
• Power users and early adopters get an insiders channel for bleeding-edge features, with crash-loop guards, version pinning, and rollback so things don't break silently.
• Teams standardizing on azd get azd update with opt-in auto-updates, version pinning, and safety guarantees so controlled rollouts are straightforward.

Goals (in scope)

• Ship the extension framework as GA: stable APIs, Go SDK, 3rd-party submission pipeline, documentation
• Ship azd ai commands as GA: full agent lifecycle from init to invoke in < 5 minutes
• Deliver first-party Foundry extension(s) built on the extension framework, proving the extensibility model
• Launch insiders channel and azd update: safe, opt-in auto-updates with rollback protection

Non-goals (out of scope)

• Building the hosted agents platform; azd is the CLI surface, not the backend
• SDKs for every gRPC language on day one; Go is first-class, others on demand
• Replacing the Foundry portal or SDK; azd is the CLI entry point
• Forced auto-updates; always opt-in, always recoverable

Success criteria

• Extension framework reaches GA with stable APIs and documentation
• azd ai commands reach GA with a complete agent lifecycle (init, deploy, invoke, down)
• Init-to-running-agent time < 5 minutes (measured end-to-end)
• First-party Foundry extension built and working on the framework
• 3rd-party extension submission pipeline operational with at least one external submission
• Insiders channel active with early adopters providing pre-release feedback
• azd update command works with opt-in auto-update and rollback/recovery

Dependencies

• Azure AI Foundry team: Platform APIs for agent hosting, model deployment, capacity resolution
• Azure Dev Tools (Pillar 2): Coordination on azd tool so updates feel cohesive
• Security/signing: Extension signing infrastructure for 3rd-party submissions