Burned 1 Billion Tokens Building an Agent Client, Got an Admin Panel Instead
sujingshen · x · 2026-09-01
The author used himself as a test subject to build a daily-use macOS personal agent client (core needs: chat + memory), letting GPT do all the coding and burning roughly 1 billion tokens in the process.
The result was a sharp contrast: instead of a chat-first assistant, GPT delivered an "admin dashboard" — 76 tests all green, yet the logs showed no model was invoked.
His takeaway: it's not that the model can't build it, but that it gravitates toward "architecturally correct" code over a product that's actually usable. Non-coders in AI-assisted development risk becoming full-time sponsors: paying, nodding, and staring at green checkmarks.
More from coding & agent
- Claude Code agent uses x402 wallet to buy equity data on-demand — kleffew94 · 2026-09-01
- Released ictexam-mcp: MCP server for exam authoring and auto-grading — ictinnovations · 2026-09-01
- Agent scammed by rug-pulled endpoint, x402 Trust offers history-based risk checks — MountainAssignment36 · 2026-09-01
- Lovable declares the 'model picker' era over, uses 1B+ prompts for auto-routing — tylerbruno05 · 2026-09-01
- Built Pattern MCP to help coding agents verify UI scope before coding — DonR954 · 2026-09-01
- Retrying LLM failures only helps if something can actually change — Available_Witness581 · 2026-09-01