Getting started with AI Agents and Multi-Agent Systems
0 likes142 views
I will guide you through the transformative journey of AI agents. As we stand on the brink of a new technological era, the session will explore the evolution of platform shifts from Client-Server to AI, highlighting the growing opportunity for autonomous AI. By 2028, it's projected that at least 15% of day-to-day work decisions will be made autonomously through agentic AI, a significant leap from 0% in 2024. We'll dive into the practical applications of AI agents, showcasing how they can enhance productivity, efficiency, and cost reduction by handling complex workflows and adapting to changing requirements. The session will cover the fundamentals of AI agents, their components, and the spectrum of their capabilities, from simple chatbots to advanced multi-agent systems. Attendees will gain insights into building AI agents, understanding their use cases, and the considerations for providing the right context, tools, and security. The session will also address the challenges of building agents from scratch and the importance of interoperability, scalability, and real-time processing. Join us to learn how AI agents can revolutionize your workflows, automate mundane tasks, and pave the way for a future where AI-driven decision-making is the norm.
![Agent
[Semi]-autonomous software that can be given
a goal and will work to achieve that goal
without you knowing in advance exactly how it's
going to do that or what steps it's going to take.](https://image.slidesharecdn.com/gettingstartedwithaiagents-maximsalnikov-250818140751-845183f6/85/Getting-started-with-AI-Agents-and-Multi-Agent-Systems-16-320.jpg)
Getting started with AI Agents and Multi-Agent Systems
- 1. Getting started with AI Agents Maxim Salnikov Senior Solution Engineer at Microsoft
- 2. I’m Maxim Salnikov Building on web platform since 90s Delivering AI & Cloud tech enablement sessions for the customers and partners at scale Speaking, training, blogging: Webdev, Cloud, Generative AI, Prompt Engineering Helping developers to succeed with the Dev Tools, Cloud & AI in Microsoft
- 4. Gen AI “Copilots” Agents ?
- 5. The opportunity for autonomous AI is growing… By 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024.* * Gartner® 2025 Top Strategic Technology Trends E-Book, Gene Alvarez and Tom Coshow, 2024, emt.gartnerweb.com/ngw/globalassets/en/information-technology/documents/trends/2025-top-tech- trends-ebook.pdf. GARTNER is a registered trademark and service mark of Gartner, 2nc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.
- 6. What is the usecase for the agentic AI?
- 7. It looks like there might be three different issues that could lead to too much noise: Unbalanced Load If the clothes inside are unevenly distributed, especially during the spin cycle, the drum can wobble and cause loud banging or thumping sounds. Worn Drum Bearings Over time, the bearings that help the drum spin smoothly can wear out, leading to grinding or rumbling noises, particularly during spinning. Loose or Damaged Parts Items like shock absorbers, suspension springs, or drive belts can become loose or damaged, resulting in clunking, squealing, or rattling sounds. Today, many customers use generative AI for knowledge retrieval (aka RAG)… Before AI Agents Search: W100 manual A customer is having an issue with their washing machine W100 making too much noise; can you help me diagnose the issue? You should create a ticket for a technician to take a look Customer Customer Support Emails, calls, or messages Searches internal knowledge base
- 8. But most downstream processes are still fragile and require meaningful human involvement Before AI Agents Customer Customer Support Emails, calls, or messages Ticket form Creates ticket Database API RPA Automation Sends notification Searches internal knowledge base
- 9. But most downstream processes are still fragile and require meaningful human involvement Before AI Agents Microsoft confidential 9 Automation cannot self- heal or handle novel cases RPA breaks whenever UI is updated or changed Employees waste time structuring data Teams must constantly monitor existing RPA scripts to ensure they do not break. Nearly all automation today starts with form over data which is slow and error prone. Complex logic is needed to handle unexpected edge cases within automation systems.
- 10. K Inefficiency and High Operational Costs Organizations face significant challenges with complex, manual, and time-consuming workflows and minimal automation Lack of Visibility & Control Human Error Scalability Issues
- 11. Autonomy and Decision making Increased Productivity Enhanced Efficiency Cost Reduction Handling complex workflows Adapting to changing requirements Collaborating with humans A new frontier: Agentic AI
- 12. LLMs update RPAs if the UI updates or changes Tool calling LLMs can self- heal or handle novel cases Here you go! Would you like me to submit the JSON to the API? With AI Agents, you can make your automation more resilient while removing the mundane work After AI Agents Microsoft confidential 12 { "name" : "John Clar "issue": ”Washing m } Convert the customer conversation into a JSON that looks like this… Transcript Click on the submit button for the scheduling form Looks like the ID of the order form is submit-order, I’ll go ahead and click it now. Submit form The form has been successfully submitted, is there anything else you need? Assign technician LLMs spend little time structuring data Update work order Let me try again with an available technician. Update work order Alex declined because they're on leave Successfully assigned Lindsay
- 13. LLMs update RPAs if the UI updates or changes Tool calling LLMs can self- heal or handle novel cases And when combined with system events, you can fully automate business processes with AI Agents After AI Agents Microsoft confidential 13 Click on the submit button for the scheduling form Assign technician LLMs spend little time structuring data Convert the customer conversation into a JSON that looks like this… Transcript Here you go! Would you like me to submit the JSON to the API? { "name" : "John Clar "issue": ”Washing m } Looks like the ID of the order form is submit-order, I’ll go ahead and click it now. Submit form The form has been successfully submitted, is there anything else you need? Update work order Let me try again with an available technician. Update work order Alex declined because they're on leave Successfully assigned Lindsay
- 14. AI Agent Use Cases • Assists employees in booking business trips • Integrates with Tripadvisor, Outlook, and SharePoint • Books via Teams chat or email • Uses OCR to gather receipts • Automates expense report submission and tracking Travel Booking & Expense Management • Personalized onboarding assistant for new hires • Uses LLMs grounded in HR data from SharePoint • Provide relevant training materials • Schedule orientations and set up software accounts • Monitor task completion and ensure efficient onboarding Employee Onboarding • Diagnoses issues by referencing history and product manuals • Provides tailored solutions or escalates through automated workflows • Creates tickets and schedules follow-ups • Updates CRM records, enhancing future support Personalized Customer Support • Analytics data from data lake and data warehouse • Responds to user requests in natural language • Generates insights, visualization, and sends via Teams or email • Automates data handling for real-time, effortless decision-making Data Analytics and Reporting
- 16. Agent [Semi]-autonomous software that can be given a goal and will work to achieve that goal without you knowing in advance exactly how it's going to do that or what steps it's going to take.
- 17. What is an AI agent? LLM Instructions (prompts) Tools Agent + + An AI agent is a micro-service that takes unstructured messages, optionally invokes other APIs and returns messages 1 2 3 Input System events User messages Agent messages 1 Tool calls Knowledge Actions Memory 2 Output Agent messages Tool results 3
- 18. Unlike previous tools, AI Agent can… reason, act and learn
- 19. Agentic patterns Acting aka tool calling Learning aka memory Reasoning aka planning I’ve captured the underlying issue in the details { "name" : "John Clar "issue": ”Washing m Create a JSON describing the user issues Check logs Machine is past due for maintenance Go beyond transcriptions and slot filling by allowing agents to deduce the why I’ll look at the logs to see what happened Transcript Fill out and submit the workorder Populate field: name The form has been successfully Agents can take a simple request and chain multiple tool calls to complete the ask Only make a mistake once. With memory, agents can recall prior experiences Populate field: issue Name field populated Issue field populated Submit form Update work order Let me try again with an available technician. Update work order Alex is already booked at this same time Successfully assigned technician Assign technician Saved memory: Alex is OOF next week
- 20. An AI agent in action… LLM Instructions Tools Agent User: My washer machine is making weird noises Agent: Recall user washing machine Tool #1 (KYC): User has w100 with sn: 142352 Agent: I’ll see what could be causing your w100 issues Agent: Se<arch manuals Tool #2 (Doc): Here is relevant parts of the manual: <…> Agent: It looks like there are three common reasons Agent: I’ll check your machine if it has any notices Agent: Check logs for 142352 Tool #3 (Log): Here is relevant log items for 142352: <…> Agent: It looks like your machine is due for maintenance and is still under warranty. Want me to schedule an appointment? User: Sure! Let’s target early next week Agent: Look for available times Tool #4 (Scheduler): Here is available time slots: <…> Agent: Would Monday at 1:00pm work? User: Let’s do it! Agent: Create workorder Tool #5 (Appointment): Order creation confirmed Agent: Done! Is there anything else I could help you with? GPT-4o You’re responsible for helping customers with their appliances Appliance manuals Check appliance logs Create workorder Recall user details
- 21. Agents range from simple to complex Tasks can vary in level of complexity and capabilities depending on your need Simple Advanced Chatbots Answer questions for users with retrieval augmented generation Copilots Performs actions on-behalf-of a user with human supervision Automated agents Automatically complete tasks as background jobs on-behalf-of users Knowledge tools + Action tools + Triggers
- 22. Spectrum of LLM-based Solutions document summary prompt answer question data query prompt + data docs conversation data query data docker databases docs history of work output input No Agent Very narrow one shot task Ex: log to JSON Single Agent Very clearly scoped iterative task Ex: providing an answer with supporting evidence to a complex question Multi-agent Systems Wide scope complex use case requiring diverse skills Ex: Propose 2 Instagram marketing campaigns including assets that would leverage the top 2 recent trends in our past quarter US Sales to boost our mailing list user base and predict the impact of each campaign VALUE
- 23. Building Agents
- 24. Components of AI Agents 1. User Experience: The prompt that starts the whole process of the agent execution, as well as the human- agent interaction. 2. Knowledge: Search and retrieve information from online sources or company knowledge base, to ground the models. 3. Actions: Helps the agents perform certain actions (e.g. send an email, write a report) through connections to key applications. 4. Memory and Threads: Captures and stores the past interactions for hyper-personalisation and increased human-like interactions 5. Autonomy: Leaves the agents perform the tasks through an event-driven, trigger-based approach. 6. Foundation Models: Enables AI Agents to think throughout the process, helping to plan and reflect. 7. Orchestrators: Whether to be the client-side code or orchestrate across multiple agents across multiple cloud, orchestrators are key to bringing everything together. Orchestrators Autonomy Foundation models 1 Knowledge Actions Memory and Threads User experience
- 25. AI Agent Considerations Providing agents with the right context Knowledge Giving agents access to the tools needed to complete tasks Actions Ensuring agents have access to only to the data and services they need Security Ensuring agents complete tasks correctly Evaluation
- 26. Building agents from scratch is difficult Tool Integration Creating a cohesive system through complex integration of various tools and APIs that have different interfaces, data formats, and requirements. Interoperability Achieving interoperability between different tools and platforms to ensure that data can be shared and understood across different systems. Scalability Handling increased data volumes, more complex computations, and higher user loads without degrading performance. Real-time Processing Ensuring tools can handle real-time requirements without significant latency. Maintenance Making labor–intensive updates to integrated tools for compatibility with new versions and prevention of obsolescence and security vulnerabilities. Flexibility Modifying or customizing existing tools or developing new ones to meet unique requirements. Error Handling Ensuring errors are handled gracefully and continue functioning despite tool failures or unexpected inputs is critical for reliability. Security Implementing robust encryption, access controls, and compliance with privacy regulations to protect sensitive data.
- 28. AppCode.class Search ( minPrice, maxPrice, category ) Agentic Pattern - Tools Calling Users What electronics do you have between $10 and $50 3 LLM Model Tool Calling 4 { tool_name: ‘product_search’ tool_args : { min_price : 10 , max_price: 50, category: ‘electronics’ } Tool Invocation Request 5 Tool Creation 1 AppCode.class @Tool( name =‘product_search‘, description =‘search for available products’ ) Search ( minPrice, maxPrice, category ) 2 Tool Binding 6 Tool Execution App
- 29. Agentic Pattern - ReAct Planning with Tools Calling Users Pay this bill for me bill-abc123.jpeg Payment Agent Bill abc123 successfully paid Planning while ( new tools execution request ) 1 scanImage ( bill-abc123.jpeg ) abc123, 100$, payeename 2 transactionHistory ( abc123 ) no results 3 paymentService (abc123, 100$,payeename) ok Instructions: • You are a home banking assistant allowing users to pay the bill uploading a picture • Always check if a bill has already been paid before submitting a payment • Confirm the payment result Tools: • scanImage ( filename ) • transactionHistory ( billId ) • paymentService ( billId, Amount, Payee)
- 30. Agentic Pattern - Memory Short Term • Access steps info in one loop iteration • Shared state context • Chat history Long Term • Access steps info in long running conversation • State persistence Conversation History Truncation • Trim by tokens • Trim by message count • Trim + summary (LLM call required) First Chat Session Second Chat Session User Chat App Memory Store Write to memory store, scoped to user_id User Chat App Memory Store Read from memory store, scoped to user_id
- 31. Agentic Pattern - Flow control Looping Termination • MaxIterations • Message termination • Human step /Human in the loop Human in the loop • Action execution approval • Escalation • Data review Payment Agent User Review Pay this bill for me Proceed Please approve payment Bill paid Tools Call Loop Run 1 Tools Call Loop Run 2 scanImage transactionService scanImage transactionService paymentService Stop Completed
- 32. A Generative AI Application with Multiple Agents “If I have any vacation days left, book 3 days for me at the start of next week.” User “I have confirmed the user has 5 days left” Data Agent “I have booked the days In the HR system” HR Agent “I have created a calendar invite” Booking Agent Knowledge Sources (search, files, databases, storage etc.) Actions (Pre-built or custom tools to automate processes) Agents perform only specific assigned tasks Agents are not overloaded with complex prompts Agents only have access to specific tools and data it needs to complete its assigned task Benefits
- 33. Typical Multi-Agent Architectures Network Maker-Checker worker reviewer Supervisor Custom Hierarchical Hand-Off
- 34. Takeaways
- 35. Learnings and next steps 1 Define and identify usecases where AI Agents can be effective 2 Explore LLM, framework, and architecture options to choose the best one for your usecase 3 Use AI Agents to build AI Agents! 4 Never stop AI upskilling journey
- 36. A free online course by Microsoft: 10 lessons teaching everything you need to know to start building AI Agents
- 37. Tusen takk! Connect with me on LinkedIn to follow the latest AI Agent advancements + request: • Link to a FREE “AI Agents for Beginners” course • Extended version on this deck (including common AI Agents patterns) • Similar (or tech/biz deeper) session for your company