3 Weeks of Logs Show 40% of Messages to Coding Agents Are Overhead
Lucky-Group9525 · reddit · 2026-09-21
The author exported and classified every message sent to his coding agents (Claude Code, Kimi, Cursor, Copilot) over 3 weeks: 2,116 genuine messages, 96/day. Only 55% was real work; 13% corrected drifting agents, 9.5% asked for progress, 6% manually relayed info between agents, 4% were "go/continue" nudges, 3% repeated rules agents never carry over. Worst day: 33 manual relays between two agents on the same experiment. Over half of "user messages" were actually his own scripts. Progress polling and cross-tool rule reuse remain unsolved.
More from coding & agent
- ττ-Bench: Best coding agent passes just 23.9% of realistic agent-building tasks — rohanpaul_ai · 2026-09-21
- ττ-bench: Best Coding-Agent Setup Passes Just 23.9% of Real Client Simulations — rohanpaul_ai · 2026-09-21
- Dev Launches Made With Jev, a Free Directory Cataloging Demos, Tools and Skills for the New Model — Sea_Supermarket_5891 · 2026-09-21
- User's Persistent Grok Agent Buys Raw Milk While They Sleep — RachelVT42 · 2026-09-21
- Which Model Actually Understands Reverse Engineering? Dev Seeks MCP Workflow for Ghidra and IDA Pro — obese_coder · 2026-09-21
- Graphify + Obsidian + Claude Code: turn file piles into a browsable knowledge graph for agents — alex_verem · 2026-09-21