{
  "id": 8748955,
  "title": "I analyzed 3 weeks of my own messages to coding agents. 40% of what I typed was not real work. Is it the same for you?",
  "url": "https://urgent.news/2026/09/20/i-analyzed-3-weeks-of-my-own-messages-to-coding-agents-40-of-what-i",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-20T19:00:35.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/torukmakto2992/i-analyzed-3-weeks-of-my-own-messages-to-coding-agents-40-of-what-i-typed-was-not-real-work-is-3e17"
  },
  "original_language": "en",
  "account": "I conducted an analysis of three weeks' worth of my own interactions with various coding agents, primarily Claude Code, Kimi, Cursor, and Copilot. To ensure accurate results, I exported each message I sent over this period and had them categorized by purpose. I didn't rely on keyword searches, as they yielded inaccurate numbers; instead, I removed automated traffic and reviewed the remaining 2,116 messages, which averaged about 96 messages per day.\n\nThe breakdown of my interactions with these agents revealed that 55% of my messages were for real work, such as new tasks, questions, and decisions. A significant portion, 40%, was categorized as overhead. This included 13% correcting the agents when they drifted from their tasks, changed models or scope without my permission, or stopped responding. Additionally, 6% of messages were about progress updates, 4% about manually transferring information between agents or chats, and 4% of messages were simple affirmative responses like \"go,\" \"yes,\" or \"continue\" due to the agents pausing for a response.\n\nOther notable categories included 3% of messages repeating rules I had previously given, and 3% covering other miscellaneous tasks like using slash commands or fragments. The most frequent cause of corrections was related to UI work, with agents requiring corrections one screenshot at a time.\n\nAcross different agents, the same rules emerged: conserve tokens, ask before changing models, and verify before releasing. Interestingly, a rule given to one agent did not apply to others, and I found myself repeating the same rule to multiple agents on various occasions. Despite the constant anxiety over token usage, agents often began large runs without clearly communicating the costs involved.\n\nSurprisingly, more than half of the \"user messages\" in my logs were not actual interactions but rather scripts and test harnesses running under my normal configuration. This raises questions about the efficiency of our current communication with coding agents. Are these same issues experienced by others, or does the overhead manifest in different ways? Which aspect costs users the most time and effort? Have there been any solutions to issues like progress polling or the rule-don't-carry-over problem, particularly when using various tools?",
  "summary": "I run several coding agents in parallel, mostly Claude Code and Kimi, sometimes Cursor and Copilot. I felt slow, so I exported every message I sent them over 3 weeks and had each one read and labeled by what it was for. I did not use keyword search. I tried that first and it gave wrong numbers. After removing automated traffic, 2,116 messages were really mine, about 96 a day. What my messages…",
  "key_points": [
    "40% of user messages were categorized as overhead tasks",
    "Common corrections related to UI work and agent behavior",
    "Half of messages were scripts and test harnesses"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}