Build AI Research Agents with LangGraph: React, Planning, and Memory
Build five agentic patterns in LangGraph, from fixed workflows to LLM agents that plan, evaluate and remember, using real SEC filings.
Build a financial research agent using five LangGraph patterns and real SEC filings. Learn to compare company performance, track financial trends, identify emerging risks, validate findings, and deliver evidence-based answers while controlling execution costs and errors.
Workshop description
- Build financial research agents using five LangGraph patterns and real SEC filings.
- Develop agents that plan, evaluate evidence, maintain memory, and control costs.
Who this workshop is for
- Data scientists, ML engineers, AI developers, Python developers, and software engineers.
- Technical leads and AI practitioners building reliable, production-focused agents.
What you will learn
- Implement fixed workflows, ReAct, reflection, planning, and four memory types.
- Evaluate agent reliability, execution traces, LLM calls, and costs.
By the end, our agent can handle questions such as
- How have company revenue and R&D spending changed over five years?
- Which financial risks emerged, and how do companies compare?
Prerequisites
- Python 3.11+, type hints, Pydantic, and basic LLM API knowledge.
- An LLM provider or Ollama, plus Docker or Neo4j. No LangGraph experience required.
Key takeaways
- Build reliable agents with validated plans, controlled retries, and traceable decisions.
- Develop reusable research workflows that deliver evidence-based answers efficiently.
Meet your Instructor
- Dr. Alessandro Negro, Senior Director, GraphAware Products at Neo4j.
- PhD in Computer Science and author of Graph-Powered Machine Learning.
Lineup
Alessandro Negro
Good to know
Highlights
- 3 hours
- Online
Refund Policy
Location
Online event
Agenda
Workflows, agents and the anatomy of a LangGraph graph (35 min)
We'll start by separating workflows, where the path is fixed in code, from agents, where the model chooses it, and see why that choice is a design decision rather than a fashion. Then we'll take a one-shot RAG pipeline and rebuild it as a LangGraph graph, naming every building block as it appears: typed state, nodes as plain functions, conditional edges, compilation and streaming, and the two state pitfalls that catch almost everyone. The result works on simple lookups and visibly fails on anything that needs more than one step, which sets up everything that follows.
ReAct: from classic tool loop to agent that tracks its own progress (40 min)
We'll build the canonical ReAct agent twice, first with the prebuilt agent and then as a hand-written loop, adding the guards real use demands: round limits, duplicate-call detection, and tool descriptions that serve as the only tool prompt. Then we'll watch it break on longer questions, as raw tool output piles up in the context, and rebuild it as separate strategy, execute and evaluate steps that keep a compact working knowledge instead of an ever-growing message history. Same tools, same questions: fewer rounds and better answers.
Agents that review and plan: reflection and plan-and-execute (55 min)
Reflection works when it is a second opinion, not a second attempt at the same prompt. We'll implement a generator and an independent reviewer that returns a typed verdict (accept, partial, retry the research, retry the answer, give up), with Python routing the verdict and enforcing retry budgets. Then we'll put two plan-and-execute designs side by side: the classic plan, execute and replan loop, and a split-call design where a single plan is validated in code, sub-questions run as parallel ReAct loops, and part of the budget is reserved so the answer and its review always get their turn. We'll compare both on a multi-year question, in coverage and in cost.