AI Coding Agents Don't Need Bigger Models. They Need Better Tools.
summary
Frontier, inference-only migrations succeed only about 30–40% of the time because huge models still read plain text files; replacing text with a Lossless Semantic Tree plus a local MCP server exposes thousands of task-shaped recipes, yields constant-time search via a triagram index, and drops token spend (an uncovered run burned 61 million tokens while a recipe-driven run changed ~700 files in minutes). Precomputed data tables, change logs, transcripts, and artifact LSTs let agents scale across ~12,000 repos with deterministic, lower-cost edits.
AI-written summary · about
chapters (12)
- 0:00Introduction and context
- 2:13Problem with text-based code interfaces
- 5:14Need for better tools and interfaces
- 8:23Introducing MCP and recipes
- 12:38Data tables and SQL for agents
- 16:32Token efficiency and challenges
- 17:59Triagram indexes for fast search
- 19:26Pre-thinking context for agents
- 22:39New systems of record
- 26:54Next-gen artifact repositories
- 28:57Building coding agent factories
- 30:42Summary and future vision
related talks
20mAI editors – how they work, and how to master them
01h 28mAI Coding Agents and how to code them
48mAI Coding Agents and How to Code Them
37mRAG, Tooling, MCP, Agents ... and What is Missing?
01h 03mAI Coding Agents and how to code them
45mNever Send a Human to do an Agent's Search
49mAI Apocalypse Survival Kit: coding agents & Tiny Language Models
11mWe Cut 94% of AI Coding Tokens With a Local Code Index