Testing AI Coding Agents: Bloated Instruction Files Waste 83% More Tokens
RunAI_Coder · reddit · 2026-08-07
The author conducted an experiment on how AI coding assistants utilize instruction files (like AGENTS.md), revealing three key findings:
- Silent Ignorance: A major harness completely ignored the standard AGENTS.md filename, only responding to its proprietary file name.
- Lean Files Save Tokens: When a lean 950-byte file was successfully read, it reduced input tokens by a third. The savings came from preventing the agent from spending extra turns excavating CI configurations.
- Bloat is Taxing: An 86KB file bloated with architecture overviews increased token usage by 83% compared to having no file at all. The agent often got confused by the noise and still wasted turns checking the configs.
The author concludes that developers should only document what the model cannot infer directly from the code, as everything else acts as a recurring tax on every API request.
More from coding & agent
- EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agent Training — Zishan Xu · 2026-08-07
- CalibForge: Automating Agent Task Generation via Adversarial Solver Calibration — AweAI-Team · 2026-08-07
- Kin: An MCP Server That Answers From a Standing Code Graph — troyjr4103 · 2026-08-07
- Reviewing the Early Academic Evolution of LLM Agent Tool-Use Harnesses — ChrisGPotts · 2026-08-07
- Kimi K3 Audits Its Own PR via Prime Agent, Passing 99 Tests — willccbb · 2026-08-07
- Replit Founder Notes They Were Training Custom LLMs Back in Early 2023 — amasad · 2026-08-07