Going Beyond Copilot with AI Agents
summary
Argues that code assistants fail because LLMs predict common tokens from open source data and lack grounding, so half of generated code is removed or rewritten within two weeks. Shows a working Python notebook using OpenAI jp4o with a REACT loop and three actions (ping, bash, web search) to discover facts (Madeira = four islands), check local packages via pip list, and run/repair code. Demonstrates a unit test generation agent that proposes tests, compiles and fixes them so generated code actually runs.
AI-written summary · about
chapters (9)
- 0:00Introduction and speaker intro
- 1:13Code assistance promise and reality
- 8:59Why current AI tools fail in enterprises
- 10:44What developers really spend time on
- 11:05Introducing AI agents
- 15:03Live demo of AI agents
- 23:13Open source tools for agents
- 24:43Zencoder product demo
- 27:54Q&A session
related talks
41mGoing beyond copilot with AI Agents
29mDemystifying AI Agents with Python Code
35mGoing Beyond Copilot with AI Agents
48mAI Coding Agents and how to code them
23mFine Tune Your AI Toolbox with GitHub Copilot, Azure AI Foundry...
40mAI Coding Agents and how to code them
29mIs coding assistant as good as we thought
48mAI Coding Agents and How to Code Them