{
  "id": 13144655,
  "title": "Why Are Coding Agents So Dumb?",
  "url": "https://urgent.news/2026/10/09/why-are-coding-agents-so-dumb",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-10-09T14:22:04.000Z",
  "source": {
    "name": "Lobsters",
    "slug": "lobsters",
    "url": "https://mtlynch.io/why-are-coding-agents-so-dumb/"
  },
  "original_language": "en",
  "account": "Why Are Coding Agents So Dumb?\n\nCoding agents, which connect large language models (LLMs) like GPT Astra, Claude Sonnet, and GLM-5.3 to codebases and computer systems, have been hyped as the next big thing in AI-assisted development. However, their performance has left many users frustrated. In February 2025, the author used a coding agent for the first time and was initially impressed by its ability to edit files directly and fix its own errors in real time. Yet, the honeymoon quickly turned sour as the agent started exhibiting frequent bugs, stopping entirely until the user restarted it.\n\nThe core problem with coding agents, according to the author, is how poorly they manage tasks. For instance, when working on an open-source web app that generates shareable links for file uploads, the agent broke the feature into 10 subtasks and completed them one by one, despite having access to multiple CPU cores for parallel processing. The agent also struggled with multitasking, often waiting for end-to-end tests to finish before starting a commit message, seemingly ignoring that a cheaper, faster model could handle the task more efficiently.\n\nMoreover, the agent fails to recognize its own capabilities. When asked about its features, it resorts to searching online, demonstrating a lack of awareness about itself. This self-awareness issue extends to the agent's task management, often assigning tasks to the smartest model without considering difficulty levels or delegating appropriately. The author argues that an LLM could easily match each subtask's difficulty to the appropriate model, eliminating the need for human babysitting.\n\nAnother issue is the agent's communication of plans, which often read like a disorganized list of low-level design decisions rather than a coherent plan. The author suggests that if a human developer provided such a plan, it would be seen as unhelpful and unnecessary. Furthermore, the agent tends to behave like a Googling employee, constantly asking questions and searching for information instead of using its own knowledge and capabilities.\n\nIn conclusion, coding agents are far from perfect and still face significant hurdles in terms of task management, multitasking, self-awareness, and effective communication of plans. Until these issues are addressed, they may remain a bottleneck in AI-assisted development, requiring human intervention to manage and optimize their performance.",
  "summary": null,
  "key_points": [],
  "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."
}