Podcast: Building a coding agent from scratch — core loop, harness design, and why harness code expires
hugobowne · x · 2026-09-15
Hugo Bowne-Anderson's Vanishing Gradients podcast episode with Nicolay Gerold (Amp Code) builds a coding agent from scratch, covering the core loop, tools, context management and harness design.
Key takeaways:
- An agent is 'an LLM, a loop, and enough tokens'; every harness component is an assumption the model can't do it alone — and expires as models improve.
- Amp removed its experimental plan mode once models reliably followed plan-without-editing instructions; dedicated file-read tools became unnecessary as stronger models used Bash to search and filter.
- After each model upgrade, ask: which workarounds can you remove, and what more can you delegate — including challenging sub-agent outputs.
- Use AI's speed for quality: try three implementations of the same idea in an hour, iterate 20 times, polish far beyond what hand-writing allowed.
Related event: Podcast Explains How to Build a Coding Agent from Scratch(2 posts)→
More from coding & agent
- Spotting an ad hidden inside a popular agent design skill — jonathan_wilke · 2026-09-15
- Savage MCP Reads Your Logged-In Chrome Sessions Without Full Browser Control — yazoodd · 2026-09-15
- Zero code: 2 days building a drivable Paris with Codex + Tripo, 386-page open-source handbook — AlchainHust · 2026-09-15
- A Fill-in-the-Blank Book Cover Prompt with Thumbnail-Ready Output — TawohAwa · 2026-09-15
- Prompt Template for High-CTR YouTube Thumbnails: One-Shot Layout Rules — TawohAwa · 2026-09-15
- Event Poster Prompt Template: Grid Layout, One Focal Point, Full Info Checklist — TawohAwa · 2026-09-15