Download as PDF, PPTX
Uploaded byDori Waldman
Introduction to AI agent development with MCP
Introduction to AI development with MCP and AI2AI
Engineeringā¦
More Related Content
Similar to Introduction to AI agent development with MCP
![[DSC Europe 24] Lana Malic - AI Agents: The Future of Autonomous Decision-Mak...](https://cdn.slidesharecdn.com/ss_thumbnails/lanamalic-aiagents-250212223710-84219c4c-thumbnail.jpg?width=600ounds&width=560&fit=bounds)
Introduction to AI agent development with MCP
- 1.
- 2. Agent Description Software programdesigned to act autonomously Interact with its environment Achieve speci fi c goals based on its own decisions They distinct from regular software because they possess greater autonomy & learning capabilities, allowing them to perform tasks on behalf of users with minimal direct intervention ~ fi re and forget
- 3. Agent has 3main parts Model Memory Tools To make model results to be consistent, set temperature=0
- 4. Memory ~ morethan session context ā¢ User: Hi my name is bob ā¢ System: Hi bob nice to meet you ā¢ User: What is my name ? ā¢ System: I dont know ā¦ Long history cost more money/token , make LLM to have confusion Short history can return wrong answers as context is partial
- 5. Tools (function call) ā¢User: What is the weather now in london ? ā¢ System: I cant say it but I can give you weather during May in general Adding weather tool - now LLM can use it to get the weather at london today Agent needs to verify it handle tools failure
- 6. Agent loop (āmax_iterationsā,default = 5) Agent dont return result right away, instead it has an āobservationā loop to verify the result before it return the fi nal answer
- 7. Multi Agents Requires: 1) crossagents shared memory: If i am booking a vacation i dont need to tell hotel & fl ight agents the vacation date and location i can tell it once and they can update each other using shared KV storage 2) Orchestration: which agent is going to handle the next task ?
- 8. Orchestration types -Sequence / Swarm / Supervisor https://www.youtube.com/watch?v=WTr6mHTw5cM https://langchain-ai.github.io/langgraph/agents/multi-agent/#swarm https://langchain-ai.github.io/langgraph/agents/multi-agent/#swarm https://google.github.io/adk-docs/agents/work fl ow-agents/
- 9. Multi agent orchestrationtypesā¦ https://pub.towardsai.net/supervisor-style-the-king-of-multi-agent-systems-139f0990ed17 https://medium.com/@siddharthc96/building-multi-agent-swarms-a-powerful-architecture-for-complex-ai-systems-37e8d7d9c0a5 https://langchain-ai.github.io/langgraph/concepts/multi_agent/#multi-agent-architectures
- 10. MCP ~ Wrapper Providea standard way to expose tools to LLM
- 11. MCP has main3 parts Tools - allow LLM to perform actions Resources - provide information to the LLM (static data like db schema) Prompt - reusable message templates https://medium.com/@mowhobuilds/mcp-server-resources-or-tools-for-exposing-get-endpoints-162d17c4c134 https://pypi.org/project/fastmcp/#resources--templates https://mail.google.com/mail/u/0/#sent/QgrcJHsTjWBSvvxSnzcKcBkbwqGqxKGtthg?projector=1 https://www.descope.com/learn/post/mcp https://www.newsletter.swirlai.com/p/everything-you-need-to-know-about
- 12.
- 13.
- 14. Dynamic Tools Generate toolsfrom con fi g fi le https://github.com/doriwal/Agent/
- 15. Multi agent -ADK Code example Agent goal is to analyze incident by getting relevant logs from elasticSearch and send them to git agent that return relevant code and send them to analyze agent that can check logs and code in order to analyze incident root cause
- 16. Agent code example Agentcan have set of tools that will be used by the model Agent save fi nal result in output_key which is KV shared memory and can be used by other agents
- 17. Multi Agent -LangGraph (low level framework) https://levelup.gitconnected.com/gentle-introduction-to-langgraph-a-step-by-step-tutorial-2b314c967d3c
- 18. Agent 2 Agent Communicationprotocol (not agent framework) to invoke āremoteā agent over http https://www.devshorts.in/p/agent2agent-a2a-protocol-explained https://towardsdatascience.com/inside-googles-agent2agent-a2a-protocol-teaching-ai-agents-to-talk-to-each-other/
- 19. Agent 2 Agent Card- service endpoint , agent capabilities , authenticaiton Task - what agent needs to do like book a fl ight , task has id and status (submitted, working , failed, complited) Communication - request-response , push noti fi cation , stream https://medium.com/@shamim_ru/google-agent-to-agent-a2a-protocol-explained-with-real-working-examples-99e362b61ba8 https://google-a2a.github.io/A2A/topics/key-concepts/
- 20.
- 21. No code ā¢ https://github.com/langchain-ai/open-agent-platform ā¢https://www.youtube.com/watch?v=kEtYJOijCBM&t=2s (n8n) ā¢ https://www.youtube.com/watch?v=Ey18PDiaAYI&t=8999s ā¢ https://www.lang fl ow.org/
- 22. AI Framework (fromhigh to low level framework) ā¢ Google ADK ā¢ https://www.youtube.com/watch?v=P4VFL9nIaIA ā¢ CrewAI ā¢ https://www.deeplearning.ai/short-courses/multi-ai-agent-systems-with-crewai/ ā¢ Aws - https://aws.amazon.com/blogs/opensource/introducing-strands-agents-an-open-source-ai-agents-sdk/ ā¢ LangGraph ā¢ https://youtu.be/jGg_1h0qzaM?si=B4UtWqrrPTsqNuxR ā¢ https://langchain-ai.github.io/langgraph/ ā¢ https://langchain-ai.github.io/langgraph/tutorials/get-started/5-customize-state/ ā¢ https://www.langchain.com/langsmith ā¢ https://academy.langchain.com/courses/intro-to-langgraph? _gl=1*ijypf1*_gcl_au*MTQ5OTgzNTg2LjE3NDg5NTE3OTE.*_ga*NjM2NzAwODc5LjE3NDg5NTE3OTE.*_ga_47WX3HKKY2*czE3NDg5NTE3OTEkbzEkZzEkdDE3NDg5NTM0ODAkajYwJGwwJGgw ā¢ Microsoft AutoGen ā¢ https://www.youtube.com/watch?v=JmjxwTEJSE8&pp=0gcJCdgAo7VqN5tD ā¢ OpenAI ā¢ https://openai.github.io/openai-agents-python/ My 2 cents: Start with google adk Its simple and you will focus on your buisness logic not of graphs
- 23. With AI Assistantwe can run faster But so can our competitors