mtlynch reviews the tooling gaps in coding agents
What happened
On October 9, Hacker News featured mtlynch's analysis of the tooling layer around coding agents. Drawing on his experience with Claude Code, Codex and OpenCode, he argues that progress in AI-assisted development comes mainly from models, while coding agents remain the bottleneck. He lists serial task execution, no model switching by difficulty, hard-to-read plans, work stopping over small problems, and weak permission boundaries. He wants OS-level sandboxes, automatic task assignment and autonomous decision-making. On why agents get little investment, he guesses managers favor demos and model benchmarks, and says that explanation does not satisfy him. That guess is his own, not a confirmed cause.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageWhy Are Coding Agents So Dumb?
mtlynch基于使用Claude Code、Codex和OpenCode等工具的经历，认为AI辅助开发的进步主要由模型推动，编码智能体仍是瓶颈。他列举了任务串行执行、缺少按难度切换模型、计划难读、因小问题停工及权限边界不足等问题，并提出操作系统级沙箱、自动任务分配和自主决策等期待。他推测企业管理者更重视演示与模型基准，导致智能体投入不足，但明确表示这一解释并不令人满意。
Heat over time
Not enough continuous observations to draw a trend yet.