Recently, we have seen a surge in requests from customers and Microsoft partners seeking guidance on building and deploying agentic solutions at various scales. With the rise of Generative AI, replacing traditional APIs with agents has become increasingly popular.

There are several approaches to building, deploying, running, and orchestrating agents on Azure. In this discussion, I will focus exclusively on Azure-specific tools, services, and methodologies, setting aside Copilot and Copilot Studio for now. This article describes the options available as of today.

1. Azure OpenAI Assistants API:
This feature within Azure OpenAI Service enables developers to create conversational agents (“assistants”) based on OpenAI models (such as GPT-3.5 and GPT-4). It supports capabilities like memory, tool/function calls, and retrieval (e.g., document search). However, Microsoft has already deprecated version 1 of the Azure OpenAI Assistants API, and version 2 remains in preview. Microsoft strongly recommends migrating all existing Assistants API-based agents to the Agent Service. Additionally, OpenAI is retiring the Assistants API and advises developers to use the modern “Response” API instead (see migration detail).

Given these developments, it is not advisable to use the Assistants API for building agents. Instead, you should use the Azure AI Agent Service, which is part of Azure AI Foundry.

2. Workflows with AI agents and models in Azure Logic Apps (Preview) – As the name suggests, this feature is currently in public preview and is only available with Logic Apps Standard, not with the consumption plan. You can enhance your workflow by integrating agentic capabilities. For example, in a visa processing workflow, decisions can be made based on priority, application type, nationality, and background checks using a knowledge base. The workflow can then route cases to the appropriate queue and prepare messages accordingly. Workflows can be implemented either as chat assistant or APIs.

If your project is workflow-dependent and you are ready to implement agents in a declarative way, this is a great option. However, there are currently limited choices for models and regional availability. For CI/CD, there is an Azure Logic Apps Standard template available for VS Code you can use.  

3. Azure AI Agent Service – Part of Azure AI Foundry, the Azure AI Agent Service allows you to provision agents declaratively from the UI. You can consume various OpenAI models (with support for non-OpenAI models coming soon) and leverage important tools or knowledge bases such as files, Azure AI Search, SharePoint, and Fabric. You can connect agents together and create hierarchical agent dependencies. SDKs are available for building agents within agent services using Python, C#, or Java. Microsoft manages the infrastructure to host and run these agents in isolated containers. The service offers role-based access control, MS Entra ID integration, and options to bring your own storage for agent states and Azure Key Vault keys. You can also incorporate different actions including invoking a Logic App instance from your agent. There is also option to trigger an agent using Logic Apps (preview). 

Microsoft recommends using Agent Service/Azure Foundry as the destination for agents, as further enhancements and investments are focused here.

4. Agent Orchestrators – There are several excellent orchestrators available, such as LlamaIndex, LangGraph, LangChain, and two from Microsoft—Semantic Kernel and AutoGen. These options are ideal if you need full control over agent creation, hosting, and orchestration. They are developer-only solutions and do not offer a UI (barring AutoGen Studio having some UI assistance). You can create complex, multi-layered agent connections. You can then host and run these agents in you choice of Azure services like AKS or Apps Service.

Additionally, you have the option to create agents using Agent Service and then orchestrate them with one of these orchestrators.

 

Choosing the Right Solution

The choice of agentic solution depends on several factors, including whether you prefer code or no-code approaches, control over the hosting platform, customer needs, scalability, maintenance, orchestration complexity, security, and cost.

Customer Need: If agents need to be part of a workflow, use AI Agents in Logic Apps; otherwise, consider other options.

No-Code: For workflow-based agents, Logic Apps is suitable; for other scenarios, Azure AI Agent Service is recommended.

Hosting and Maintenance: If Logic Apps is not an option and you prefer not to maintain your own environment, use Azure AI Agent Service. Otherwise, consider custom agent orchestrators like Semantic Kernel or AutoGen to build the agent and services like AKS or Apps Service to host those.

Orchestration Complexity: For simple hierarchical agent connections, Azure AI Agent Service is good choice. For complex orchestration, use an agent orchestrator.

Versioning - If you are concerned about versioning to ensure solid CI/CD regime then you may have to chose Agent Orchestrators. Agent Service still miss this feature clarity. We have some work-around but it is not robust implementation. Hopefully we will catch up soon with a better versioning solution.

 

Summary:

When selecting the right agentic solution on Azure, consider the latest recommendations and platform developments. For most scenarios, Microsoft advises using the Azure AI Agent Service within Azure Foundry, as it is the focus of ongoing enhancements and support. For workflow-driven projects, Azure Logic Apps with agentic capabilities may be suitable, while advanced users can leverage orchestrators for custom agent architectures

Back at it today after a three day weekend to Texas. Lots of AI content today!

[article] How do AI-assisted code fixes impact code review time? Don’t expect a “drop-in efficiency boost” as noted here. But Meta’s research did highlight a few good pieces of advice.

[blog] Writing effective tools for agents — with agents. Great advice in this Anthropic post. Read it to learn more about tools and practices for building effective ones.

[blog] Gemini CLI Tutorial Series — Part 11: Gemini CLI Extensions. The extensibility of the Gemini CLI will spark another monster wave of creative use cases for agentic CLIs.

[article] Agents show promise, but widespread usage in the enterprise remains elusive. Two things can be true. It’s true that there’s excess hype and a dearth of wild success stories with AI. It can also be true that our industry hasn’t skilled up yet and the tech is improving rapidly.

[blog] How AI made Meet’s language translation possible. Case in point. Real-time language translation will be ubiquitous in a short time.

[blog] GPT‑5-Codex and upgrades to Codex. Looks like a good upgrade to OpenAI’s coding models and tools. I can’t track all the names, though. More here.

[blog] Let AI Interact with Your App via MCP. How long until agentic users outnumber human users for your app? If you’re a SaaS provider, it’s going to happen more quickly than you think.

[blog] Listen to a discussion on how AI can power scientific breakthroughs. The commercial work we do matters because it makes it possible to fund such amazing work for the benefit of humanity.

[docs] Serve open LLMs on GKE with a pre-configured architecture. I think most of us want smart infrastructure default values that we can trust and accept. Here’s an example of what you can do when you have that.

[article] At 20, Techmeme has never been hotter. I’ve been a reader of Techmeme for years, as it’s the best source I know of to get a “what’s going on in tech today” view of our industry.

[blog] OpenTelemetry Protocol comes to Google Cloud Observability. Looks like we’re picking OpenTelemetry as the winner, and it’s getting baked in for telemetry across Google Cloud.

[article] AI developer certifications tech companies want. I can’t imagine anyone getting hired because they’re certified, but it certainly signals a level of effort and interest in a topic.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:

Join Dale Dougherty and a group of exhibitors for an inside look at what’s to come at Maker Faire Bay Area 2025!

The post A Sneak Peek at Maker Faire Bay Area 2025 appeared first on Make: DIY Projects and Ideas for Makers.

The Model Context Protocol (MCP) ecosystem is enhancing AI development with a public registry for server discovery and a secure gateway for agent interactions. This initiative, featuring the recently launched MCP Registry and the Linux Foundation's Agentgateway project, streamlines the management of AI tools, fostering collaboration and security for engineering teams.

By Andrew Hoblitzell

404 Media: A data broker owned by the country's major airlines, including American Airlines, United and Delta, is selling access to five billion plane ticketing records to the government for warrantless searching and monitoring of peoples' movements, including by the FBI, Secret Service, ICE, and many other agencies, according to a new contract and other records reviewed by 404 Media. The contract provides new insight into the scale of the sale of passengers' data by the Airlines Reporting Corporation (ARC), the airlines-owned data broker. The contract shows ARC's data includes information related to more than 270 carriers and is sourced through more than 12,800 travel agencies. ARC has previously told the government to not reveal to the public where this passenger data came from, which includes peoples' names, full flight itineraries, and financial details. "Americans' privacy rights shouldn't depend on whether they bought their tickets directly from the airline or via a travel agency. ARC's sale of data to U.S. government agencies is yet another example of why Congress needs to close the data broker loophole by passing my bipartisan bill, the Fourth Amendment Is Not For Sale Act," Senator Ron Wyden told 404 Media in a statement.

Read more of this story at Slashdot.

An unlisted and now-removed video from Meta showed off a new pair of Ray-Ban branded smart glasses with a display and a wristband to help control them, as reported by UploadVR. The video also features a new pair of wraparound “Sphaera” Oakley smart glasses with a camera on the nose, meaning that Meta itself may have inadvertently leaked two of its biggest announcements from its Connect event this week.

For the glasses with a display, the video includes branding that says “Meta | Ray-Ban” and “Display,” so perhaps Meta will call them the Meta Ray-Ban Display. Based on a brief clip included in UploadVR’s article, the actual display is in the right lens of the glasses, and the clip shows how the glasses will let you do things like ask Meta AI a question, look at a map on the glasses to help you navigate, translate a sign, or use your hand with the wristband to “write” a reply to a chat.

Earlier this year, CNBC reported that the wristband for Meta’s codenamed “Hypernova” glasses with a display would use surface electromyography (sEMG) technology to interpret signals from hand movements to help you control the device, so the clip is presumably showing that tech in action.

Meta CEO Mark Zuckerberg will be hosting a Connect keynote at 8PM ET on Wednesday, where it seems like these new glasses will make their official debut. The Verge will be covering the keynote live.

Correction, September 15th: An earlier version of this article speculated about translucent versions of the HSTN Oakley glasses, but those are already on sale.