Token Efficiency and Practical Use of AI Coding Agents
코딩/개발 | Mon Jul 13 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources
Systima compared token overhead between Claude Code and OpenCode, Ploy migrated to GPT-5.6, Terry Tao ported Java applets via AI agents, and Doug Turnbull argued for writing code by hand.
Analysis
[Systima] measured token overhead comparison between Claude Code and OpenCode [1]
- Claude Code sends 33k tokens before the prompt
- while OpenCode sends 7k tokens
- Approximately 4.7x difference on Sonnet 4.5
- 3.3x on Fable 5
- Claude Code triggers up to 54x more cache writes
- Config files like AGENTS.md add 20k tokens per request
- Subagent usage increases tokens from 121k to 513k
[Ploy] migrated a production AI agent to GPT-5.6 Sol [2]
- 2.2x faster and 27% cheaper than Claude Opus 4.8
- The first model to beat Opus in four months
- One-third of eval failures turned out to be harness assumption issues
- Required sequential fixes to tool schemas
- caching
- and reasoning replay
- Provider-specific behaviors exist even when using Vercel AI SDK
[Terry Tao] ported 1999 Java applets to JavaScript using an AI coding agent [3]
- Converted around 20 old applets to modern JavaScript in a few hours
- Added graphics upgrades such as color to the Besicovitch set applet
- Found 1 minor bug during porting and 2 bugs in the original
- Completed a special relativity visualization app abandoned in 1999 via vibe coding
- Realized an idea for a Minkowski space version of Inkscape
[Doug Turnbull] argued why developers should still write code by hand in the AI era [4]
- A software engineer's role is to build the 'factory
- ' not the software
- AGENTS.md
- skills
- and automated evals establish infrastructure for agent success
- English is an under-specified language for expressing computation
- Writing code is essential for understanding architecture and detecting fragility
- Without hand-written code
- slop accumulates and harms even the agents