Job bots expose MiniMax, Qwen limits

This is less a model problem than an orchestration problem. LLMs can reason about the task, but reliable job-application automation needs deterministic browser control, state handling, retries, and human approval at the final send step.

• –Email and LinkedIn are brittle targets: dynamic UIs, anti-bot defenses, and frequent layout changes make fully autonomous flows fragile
• –The programming language matters less than the execution layer; Playwright/Selenium-style control and structured tool calls matter more than Flutter vs Python vs C++
• –Models often look much better when used interactively than when asked to self-generate the automation code they’ll later depend on
• –For anything involving accounts, approvals, or platform ToS risk, a human-in-the-loop workflow is safer and usually more dependable
• –The post is a useful reality check for agent builders: “can chat about the task” is still far from “can consistently do the task”