Azure AI Agents are in the Public Preview – Vesa Nopanen – Future Work today

AI brokers have gotten a key enabler for companies trying to streamline processes, automate repetitive duties, and empower their staff to work extra effectively. In Microsoft 365 Copilot, we’ve already seen a number of options that target bettering productiveness for people. But, the potential for AI-driven automation goes a lot additional when you may join clever , pure language, brokers on to your personal enterprise information and processes—enter Azure AI Agent Service in Azure AI Foundry.

On this submit, you examine why AI Brokers are invaluable, and the way Azure AI Agent Service makes it simple to construct and customise these brokers. I additionally did some testing and share these experiences, together with the Code Interpreter characteristic for information evaluation. Lastly, in the long run you may learn by means of some fast tips about how one can get began.

1. What Are AI Agents and Why Should Businesses Care?
   1. Stories of Transforming Business Processes
   2. AI Agents
2. Azure AI Agent Service is now in Public Preview
3. My Testing Experience with Code Interpreter and Knowledge Sources
   1. Code Interpreter
   2. External Knowledge & RAG Testing
   3. Agent Configuration Made Simple
   4. Model Tuning Options and Limitations
4. Getting Started: The Quickstart
   1. Prerequisites
   2. Creating and Testing Your Agent
   3. Expand with SDK or Additional Tools
5. Conclusion
   1. Read more from these sources:
6. About writing this article

At its easiest, an AI agent is a self-contained “microservice” powered by a big language mannequin (LLM) or comparable AI mannequin. It’s designed to reply questions, carry out actions, and in the end automate or increase particular duties. Let’s check out some examples first.

Tales of Remodeling Enterprise Processes

A Fictional Take a look at a Multi-Agent Gross sales Ecosystem
Think about a big electronics retailer that operates in dozens of nations. They’ve separate specialised AI brokers for various points of the gross sales cycle. One AI agent handles lead qualification by scanning incoming inquiries and extracting key details about potential shoppers. Concurrently, one other agent is chargeable for product suggestions primarily based on real-time pricing and stock information. As soon as a lead is certified, a scheduling agent units up demos with gross sales reps and routinely books a gathering in Microsoft Groups, full with related paperwork connected.

These brokers may collaborate, passing data to 1 one other about lead standing or one of the best product bundles for a specific area. Due to this multi-agent method, a gross sales rep can bounce in solely when human interplay is really wanted, relatively than juggling repetitive duties like re-checking inventory or manually coordinating conferences. Your entire course of is a well-orchestrated system that frees staff from repetitive admin and information entry. The outcome: extra customized interactions with prospects, increased gross sales velocity, and a drastically decreased probability of human error.

Fujitsu: RAG and Sales Efficiency in the Real World
Whereas the above situation is hypothetical, real-life companies already leverage Azure AI Agent Service to revolutionize their gross sales processes. For instance, Fujitsu leveraged Azure AI Agent Service and Semantic Kernel to construct an AI-powered automation resolution to streamline proposal creation, enabling gross sales groups to give attention to high-value buyer engagement. The AI agent dynamically retrieves and synthesizes information from dispersed sources, making certain accuracy and relevance whereas integrating seamlessly into Fujitsu’s current Microsoft ecosystem. “We’re utilizing Microsoft’s Semantic Kernel and Azure AI Agent Service to orchestrate a number of specialised AI brokers and an orchestrator AI to coordinate them to reply questions as a workforce,” 

Cineplex: Transforming Customer Service Through Automation
Cineplex, a number one Canadian media and leisure firm, has reworked its customer support operations utilizing AI-powered automation. One in all its greatest challenges was dealing with refund requests, a time-consuming course of that took 5–quarter-hour per request. To unravel this, Cineplex applied an AI Copilot agent utilizing Microsoft Energy Platform and Azure AI. Now, visitor service brokers merely enter a reserving ID and date, and the AI handles the remaining—retrieving information, validating the request, and finishing the refund in about 30 seconds This highlights how AI brokers can optimize routine duties, increase effectivity, and improve customer support—with out changing human interplay. For companies trying to scale help operations, AI-powered automation generally is a game-changer.

AI Brokers

In contrast to fundamental chatbots, AI brokers can incorporate context from historic conversations and connect with exterior programs, permitting them to:

• Search your organization’s data base or the online.
• Course of and interpret information or real-time information.
• Make calculations, generate experiences, even run code.
• Carry out advanced duties that save staff effort and time.
• Combine with different programs, inside and exterior
• Interact different brokers within the course of: multi-agent programs
• Cope so much higher with numerous conditions higher than conventional automation. This is because of LLM of their “core” giving understanding of the objective and what’s wanted. AI Brokers are versatile and might adapt to conditions – and in addition know when to ask assist from an one other agent or an actual individual.

In different phrases, AI brokers can complement human staff by taking up repetitive or time-consuming jobs. That may imply a buyer help agent that routinely retrieves solutions from a data base, an inside finance agent that crunches finances information from spreadsheets, or a gross sales agent that triggers e-mail workflows.

From a enterprise standpoint, AI brokers have tangible advantages:

• Accelerated decision-making: Dynamic, context-aware AI reduces handbook analysis.
• Scalability: Brokers can work across the clock, dealing with duties for a number of groups concurrently.
• Consistency: They apply data and logic in a uniform manner—fewer errors as a result of human error. At this time is the time once we experiment with brokers, and it must be realized brokers may make mistakes – generally even a lot.. The objective is within the consistency and arising with new concepts the place AI can rework the method, and these received’t be reached with out experimenting, arising with difficult use circumstances and braveness to check out one thing new.

Azure AI Agent Service, now accessible as public preview within the Azure AI Foundry portal, gives a managed setting to construct, debug, and deploy these AI brokers. It’s designed in order that builders and tech-savvy enterprise customers can rapidly form an agent’s capabilities with out having to assemble all of the underlying code or infrastructure themselves. This accelerates pro-code agent improvement and is but an one other instance of fusion groups the place enterprise and builders work collectively.

Key capabilities embody:

• Prepared-to-Go Instruments & Integrations:
  • Code Interpreter: Permits brokers to execute Python code inside a safe sandbox—nice for number-crunching, information evaluation, or producing graphs.
  • Bing Search & Azure AI Search: Brokers can pull in exterior data from the online or your personal information, including related context to duties.
  • Azure Features Assist (SDK-based): Builders can expose customized enterprise logic or exterior APIs to the agent, letting it set off real-world actions.
• Conversational Reminiscence:
  Brokers can keep a thread of dialog, bear in mind particulars, and proceed the place you (or the agent itself) left off. That is dealt with securely on the server facet.
• A number of Mannequin Choices:
  Though Microsoft’s GPT-4o is a well-liked alternative, you may as well deploy different associate fashions like Cohere or Mistral within the Azure AI Foundry. (Observe: Mistral-large-2407 is turning into legacy and will not be accessible for much longer.)
• Fundamental vs. Customary Setup:
  • Fundamental Setup (supported within the Azure AI Foundry portal as we speak): You depend on Microsoft-managed assets for storage and search. Fast to begin, minimal overhead, but it surely provides much less management.
  • Customary Setup (Bicep template–solely): You deliver your personal assets (like Azure Storage and Azure AI Search) for full visibility and value administration.

I’ve spent a while exploring the brand new Brokers UI in Azure AI Foundry, placing these options to the check. Right here’s what stood out for me:

Code Interpreter

A enjoyable (although fictional) situation was exploring the terminal velocity of a laptop computer falling from an airplane. With the agent’s Code Interpreter instrument enabled, I might ask the agent to run physics-related calculations. It will possibly generate fast math scripts in Python—and that is only a easy instance concerning the Code Interpreter.

I uploaded an Excel file for the 2023 finances of the town of Vantaa (accessible as open information) to the Code Interpreter. The agent then learn and interpreted the file, making it easy to investigate finances figures, collect insights, and visualize the info.

Exterior Information & RAG Testing

I additionally examined a situation utilizing fundamental RAG (retrieval augmented technology). By importing some demo paperwork, the agent was capable of pull focused details from my very own content material, weaving them into its solutions. The RAG with AI isn’t something new anymore, however the Assistants API working behind the hood isn’t an on a regular basis instrument but. So it made sense to mess around to see the way it performs – and it was similar to I anticipated.

Agent Configuration Made Easy

The visible circulate within the Azure AI Foundry UI is deceptively easy: outline your agent’s title, add Information sources (information or indexes), and specify which Actions (instruments) the agent might use.

Presently, the one motion accessible from the UI is Code Interpreter. If you wish to combine your personal, corresponding to Azure Functions, you can do more via the SDK.

Mannequin Tuning Choices and Limitations

Fundamental tuning for Temperature and High P is well accessible within the UI, so you may alter how artistic or deterministic your agent’s solutions needs to be.

Within the UI, solely Code Interpreter is displayed as an “motion,” however the underlying Assistants API definition is versatile—new actions or customized instruments could be added as soon as they’re enabled.

I experimented with GPT-4o, which labored seamlessly. The service additionally guarantees help for non-OpenAI fashions like Cohere and Mistral, although my free Azure subscription didn’t enable me to deploy them.

General, these assessments spotlight how rapidly you may piece collectively a specialised AI agent that’s distinctive to your model, workforce, or venture. With a number of traces of code or a number of clicks within the UI, you may rework a easy chat mannequin right into a mini-assistant with actual enterprise worth.

Need to attempt it your self? Right here’s a brief information primarily based on the official quickstart.

Conditions

• An Azure subscription (create a free trial if wanted).
• The Azure AI Developer position assigned. This provides you the precise permissions to create and handle AI brokers.
• Fundamental Setup by way of the Azure AI Foundry Portal: As a result of the Foundry portal solely helps the “fundamental setup,” you’ll be utilizing Microsoft-managed storage and search behind the scenes. This will get you getting began quick.

Creating and Testing Your Agent

• Navigate to Brokers within the Azure AI Foundry portal and choose “New agent.”
• Present a reputation and add directions (e.g., “You’re a enterprise analyst specializing in forecasting.”).
  A tip: use Chat Playground’s Generate immediate characteristic to construct directions for the agent.
• Beneath “Information & Motion,” add Code Interpreter if you’d like the agent to deal with information evaluation or code execution. You can too connect as much as 20 information that your agent can learn and use for producing outputs.
• After configuring your agent, swap to the “Playground” to start chatting.
• You possibly can revise directions, tweak mannequin parameters (Temperature, High P), or add new data information and instruments.
• Affirm that your agent is responding as anticipated and refine your directions or information sources if wanted.

To include your personal Azure Features or exterior APIs, you’ll have to outline them as instruments by way of the Azure AI Foundry SDK or the Azure OpenAI SDK. That is notably useful for extra advanced automations the place the agent may, for example, replace a CRM file or ship an e-mail in your behalf.

Azure AI Agent Service is a promising step ahead in automating various enterprise processes—from information evaluation and RAG queries to extra action-oriented duties like connecting to exterior APIs. The mixture of huge language fashions, built-in instruments, and easy setup within the Azure AI Foundry UI makes it a compelling alternative for making an attempt out quite a lot of automation situations. Sooner or later (close to, I hope) we will additionally add multi-agent programs to this.

For enterprise choice makers, one key issue is how rapidly and securely it’s doable obtain operational advantages (and ROI). Whether or not you’re in finance, manufacturing, retail, or past, AI brokers supply a brand new method to faucet into supercharging enterprise processes. Suppose scaling processes that historically depend upon human intervention, to agent-driven that enhance productiveness, scale back handbook errors, and releasing tedious work (and valuable) time from people. After I discuss with prospects about Microsoft 365 Copilot, it already helps many to finish extra duties sooner than earlier than. For a lot of of them, which means much less lengthy days turning evenings – or that pile of to do duties stays in management.

Should you’re curious, I like to recommend checking the quickstart, spinning up a fundamental agent, and giving Azure AI Service UI and particularly the Code Interpreter with Assistants API a attempt. From information crunching to serving to your gross sales or help groups, you may see how briskly you may construct an pro-code agent core able to significant work.

Learn extra from these sources:

Sure, I used once more the Azure OpenAI Service reasoning mannequin o1 to assist me out with this. I offered the mannequin an extended immediate, that included my objective, insights, data of what I did and what I wished to precise within the submit. Together with the background data from Microsoft Study and articles. After that I used some prompts to refine the outcome and added instance use circumstances. Lastly I coped the textual content to the weblog and went by means of this – making use of modifications, deleting components and including new insights, and naturally photos. This sped up the precise weblog writing course of quite a bit, but it surely nonetheless took a number of of hours in whole.

Maybe for one weblog submit I’ll create a Groups assembly, that I file and transcribe, after I testing out new characteristic. Utilizing that might present fairly an distinctive base for the submit draft, that I generate with the assistance of o1. That may not be so structured as writing my chosen insights, however would positively be a distinct manner. Will or not it’s sooner? That I can discover out by testing it out.