Simonjj explores the public preview of Compose for Agents on Azure Container Apps, showing how developers can leverage serverless GPUs, Docker Compose, and a wide selection of AI frameworks to build and scale agentic applications on Azure.

Compose for Agents on Azure Container Apps and Serverless GPU (Public Preview)

Author: simonjj

Overview

Azure Container Apps (ACA) now offers first-class support for agentic AI applications via Docker’s Compose for Agents, currently in public preview. This empowers developers to easily define, deploy, and operate complex agent-based workloads—such as virtual assistants and workflow agents—using familiar Docker Compose tooling and Azure’s fully managed, serverless infrastructure.

Key Challenges in Agentic Application Development

Modern intelligent applications require weaving together LLMs, vector databases, MCP (Model Context Protocol) tools, and orchestration logic. Traditionally, this brings several challenges:

Tooling sprawl: Multiple SDKs and frameworks, tedious dependency management
Specialized hardware: GPUs are required for performant model inference and orchestration, which are costly to manage
Operational complexity: Scaling and securing multi-service applications is traditionally a full-time effort

How Azure Container Apps Simplifies Agentic AI Solutions

Azure Container Apps addresses these pain points with:

Serverless GPUs with per-second billing: Run workloads on demand, paying only for actual GPU usage—ideal for sporadic or highly variable compute needs
Sandboxed dynamic sessions: Secure, short-lived environments for running user-provided or third-party code, separating evaluation scripts or plugins from critical services
Automatic scaling and managed operations: Container Apps handle service discovery, ingress, scaling, rolling updates, and revision management—no need to maintain your own orchestrator
First-class Docker Compose integration: Define your stack in a standard compose.yaml, inclusive of models, agents, open tools, and microservices, for seamless deployment to both local and cloud environments
Integrated Model Runner & MCP Gateway: Model Runner lets users pull and expose open-weight language models via OpenAI-compatible endpoints. MCP Gateway connects agents to a curated suite of tools, enabling retrieval-augmented generation (RAG), vector search, and specialized tool invocation directly within the Compose stack

Developer Workflow and Supported Frameworks

With Compose for Agents in ACA, developers can:

Describe agent stacks declaratively: Use a single YAML file to orchestrate models (e.g., via Model Runner and MCP Gateway), agents, and services using frameworks like LangGraph, Embabel, Vercel AI SDK, Spring AI, CrewAI, and Google ADK
Maintain consistency across environments: The same compose file can launch locally or in production, enabling fast iteration and cloud readiness
Scale seamlessly: Virtually unlimited compute elasticity with serverless GPU integration and scaling

Supported Architectures and Tools

Framework agnosticism: Choose from LangGraph, CrewAI, Spring AI, or other preferred stacks
Tool integration: Leverage over a hundred ready-made tools and services from Docker’s MCP catalog for retrieval, summarization, data access, and more

Getting Started

To try Compose for Agents preview on ACA:

Install the latest Azure Container Apps Extension
Define your agentic application and dependencies in compose.yaml
Deploy with az containerapp compose up—GPU allocation, dynamic sessions, and scaling are handled automatically
Iterate locally with docker compose up, then use the same configuration file to run in Azure

For detailed instructions, visit https://aka.ms/aca/compose-for-agents

Collaboration with Docker

This innovation is the result of joint work between Microsoft and Docker, bringing together Docker Compose’s developer-favorite workflow with Azure’s robust, scalable serverless infrastructure for modern AI workloads.

Published by simonjj, November 18, 2025. For more content like this, follow the Apps on Azure Blog.

This post appeared first on “Microsoft Tech Community”. Read the entire article here