Agentic Rag: Building a coding agent (no frameworks) 🦄 ep #28

Watch Vaibhav build a coding agent from scratch, showing the differences between traditional RAG and Agentic RAG systems. This live coding session demonstrates why the hardest part isn't the agent loop - it's the tool implementation details that make or break your system.

What You'll Learn:
• The real difference between deterministic RAG and agent-driven context assembly
• Why relative paths vs absolute paths can make or break your grep tool
• How to build an agent with 16 tools (glob, grep, read, bash, web search, and more)
• Critical context engineering tricks that save thousands of tokens
• When you should (and shouldn't) use Agentic RAG in production
• Why building from first principles beats using frameworks for learning

Key Insights:
✅ 70% of the code was AI-generated, but the crucial 30% required deep understanding
✅ Most time was spent on UI/debugging tools, not the agent logic
✅ Tool implementation details matter more than perfect prompts
✅ Small optimizations (20 tokens saved × 30 calls) compound dramatically

Code & Resources:
🔗 Full source code: https://github.com/ai-that-works/ai-t...
🔗 BAML Language: https://github.com/BoundaryML/baml
🔗 Discord: https://boundaryml.com/discord


Timestamps:
00:00 Introduction to Agentic RAG Systems
02:52 Demo of the Coding Agent
05:44 Understanding Agentic RAG vs Traditional RAG
08:34 Building the Agentic RAG System
11:38 Iterative Development and Testing
14:14 Challenges in Tool Implementation
17:13 Evaluating Tool Effectiveness
19:59 Designing the User Interface for Agents
22:56 Managing State and Context in Agents
26:05 Final Thoughts and Future Improvements
31:07 Navigating Directory Structures in Code
34:16 Optimizing Grep and Read Tools
35:19 Understanding Retrieval-Augmented Generation (RAG)
39:45 Implementing Web Search and Context Efficiency
43:28 Enhancing Tool Interactions and Error Handling
49:13 Iterating on Tool Design and User Experience
55:42 Building Agentic RAG Systems
59:32 Exploring Model Performance and Cost Efficiency
01:00:54 Understanding Tool Responses and Context Engineering
01:01:30 Implementing Feedback Loops in AI Models
01:04:52 Balancing Speed and Accuracy in AI Systems
01:06:46 Designing User Interfaces for AI Interactions
01:08:53 Debugging and Improving AI Pipelines
01:13:24 Dynamic Context Management in AI Systems
01:18:19 Final Thoughts on Building Effective AI Solutions