CONFIDENTIAL
Building Agentic Systems
By Nicholas Roze-Freitas
My Background
• 10 Months in SAP Academy for Product & Engineering
• Studied Data Science and Math at UCSB
• Data Scientist building agents at SAP Concur
Agentic Systems?
• Different categories of Agentic Systems
• independent, fully autonomous
• More prescriptive implementation that
follows a flow
• LangChain’s blog writes often times Agentic
Systems in production are a combination
Complex
Reasoning
Dynamic
Decision-making
Unstructured
Data
When are Agents good to use?
Main Components Of An Agent
• LLM
• Tools
• Memory
• Short term
• Long term
Langgraph
Building Agents
• Tons of frameworks
• Langchain / Langgraph
• CrewAI
• Google ADK
• OpenAI Agents SDK
• … and more!
• Agents themselves are pretty simple
• Most frameworks implement the same loop
Also Langgraph
The Loop
• Feed Messages to LLM
• If LLM called tools
• Call tools & add results to messages
• If LLM responds with text
• Get user feedback if applicable
• Or do something custom
Again Langgraph
Some Hidden Challenges
• Security
• Malformed & incorrect tool calling
• Message history and context management
Agents
Sometimes
Hidden Challenge: Security
• LLMs can generate malicious outputs
• User/company data isolation
• Console commands
• Solution?
• Design interactions with implicit guardrails
• User/company based instances of tools
Agents
Sometimes
(again)
Hidden Challenge: Malformed Function Calls
• LLMs can fail to generate function calls
• Malformed function calling can break up
flows
• Solution?
• Define tools with primitives
• Implement custom error handling
Hidden Challenge: Context
• Agents can forget things
• Context helps agents understand their task and
get information
• Solution?
• Keep most important info at top of context
• Design RAG systems to pull context
dynamically
Agents Thinking
Really Hard
My Library of choice: LangChain & LangGraph
Why I like Lang(Chain/Graph)
• Offers high and low level APIs
• Represents agentic systems as a graph
• Low level APIs give the ability to build highly
customizable solutions
• Building blocks
• Mitigations
• Compliance
Some Companies Using
LangGraph
Getting Started
• LangGraph/Langchain offers high level APIs that are
easy to start with
• create_agent
• middleware
• Or get into the weeds and just start with LangGraph!
What I’m Excited About
• Fine-tuned small language models
• Large model inference is expensive & unnecessary
• Agent’s often handle specialized tasks
Thank you + Questions
Connect with me on LinkedIn @ Nicholas
Roze-Freitas

SAP Building Agentic Systems (Milvus Community Meetup)

  • 1.
  • 2.
    My Background • 10Months in SAP Academy for Product & Engineering • Studied Data Science and Math at UCSB • Data Scientist building agents at SAP Concur
  • 3.
    Agentic Systems? • Differentcategories of Agentic Systems • independent, fully autonomous • More prescriptive implementation that follows a flow • LangChain’s blog writes often times Agentic Systems in production are a combination
  • 4.
  • 5.
    Main Components OfAn Agent • LLM • Tools • Memory • Short term • Long term Langgraph
  • 6.
    Building Agents • Tonsof frameworks • Langchain / Langgraph • CrewAI • Google ADK • OpenAI Agents SDK • … and more! • Agents themselves are pretty simple • Most frameworks implement the same loop Also Langgraph
  • 7.
    The Loop • FeedMessages to LLM • If LLM called tools • Call tools & add results to messages • If LLM responds with text • Get user feedback if applicable • Or do something custom Again Langgraph
  • 8.
    Some Hidden Challenges •Security • Malformed & incorrect tool calling • Message history and context management Agents Sometimes
  • 9.
    Hidden Challenge: Security •LLMs can generate malicious outputs • User/company data isolation • Console commands • Solution? • Design interactions with implicit guardrails • User/company based instances of tools Agents Sometimes (again)
  • 10.
    Hidden Challenge: MalformedFunction Calls • LLMs can fail to generate function calls • Malformed function calling can break up flows • Solution? • Define tools with primitives • Implement custom error handling
  • 11.
    Hidden Challenge: Context •Agents can forget things • Context helps agents understand their task and get information • Solution? • Keep most important info at top of context • Design RAG systems to pull context dynamically Agents Thinking Really Hard
  • 12.
    My Library ofchoice: LangChain & LangGraph
  • 13.
    Why I likeLang(Chain/Graph) • Offers high and low level APIs • Represents agentic systems as a graph • Low level APIs give the ability to build highly customizable solutions • Building blocks • Mitigations • Compliance Some Companies Using LangGraph
  • 14.
    Getting Started • LangGraph/Langchainoffers high level APIs that are easy to start with • create_agent • middleware • Or get into the weeds and just start with LangGraph!
  • 15.
    What I’m ExcitedAbout • Fine-tuned small language models • Large model inference is expensive & unnecessary • Agent’s often handle specialized tasks
  • 16.
    Thank you +Questions Connect with me on LinkedIn @ Nicholas Roze-Freitas