Building LLM-based Agents
summary
Claims a reproducible pattern: an LLM (GPT‑4) acting as a decision loop that picks real tools, runs them, and composes answers. Implements Haystack pipelines as concrete tools — pycon QA tool, speaker search, web QA pipeline — wired into a zero‑shot REACT prompt so the agent iterates: choose tool, run, observe, repeat. Demo extracts a talk date from event data then queries web search for weather, showing the approach actually works end‑to‑end with retrieval‑augmented generation.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 1:25Human info search history
- 2:48What large language models can do
- 4:13Introducing agents
- 8:10Prompt engineering importance
- 8:51Haystack framework
- 11:38Haystack pipeline limits
- 12:01PyCon QA pipeline
- 13:43Prompt templates in Haystack
- 17:36RAG QA demo
- 18:38Agent definition and features
- 20:24Zero-shot REACT prompt
- 21:48Agents vs pipelines
- 22:56PyCon agent build
- 24:19Adding agent tools
- 26:08PyCon agent demo
- 26:49Summary and resources
- 27:55Q&A session part one
- 30:12Q&A session part two
- 31:59Q&A session part three
- 33:10Q&A session part four
- 34:58Closing remarks
related talks
41mBuilding LLM Agents – Challenges and Learnings
01h 43mBuilding LLM Agents (Hackathon)
09mBuilding LLM Agents with Tool Use
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
40mGo Beyond Basic RAG with Agentic Behavior
47mBuilding and Testing AI-Agent Powered LLM Applications: A Live Demonstration
01h 00mFrom LLM Theory to Practical Agentic Implementations
30mBuilding Robust, Secure LLM and Agentic AI Workflows