What This Is
A developer community recently ran a head-to-head comparison of three AI coding assistants — Cline, Kilo, and Qwen Code (tools that let AI autonomously write code, debug, and execute commands). The investigation zeroed in on "long-task context management" — that is, how an AI keeps track of where it is and what comes next after two or three hours of continuous work. The conclusion is blunt: 90% of "AI coding gone off the rails" incidents are not a model intelligence problem. They are a crude memory architecture problem.
All three approaches have flaws. Cline writes task state to a Markdown file outside the conversation and periodically re-injects it, but the file and what the model actually sees sometimes drift out of sync. Kilo is the weakest of the three: its TODO state lives directly inside the conversation history, so once the conversation is compressed (condensed into a short summary to save space) it turns into garbled text, forcing the AI to repeatedly re-read source code to recover its bearings — often trapped in an endless read-analyze-compress loop. Qwen Code stores TODO state in a fully separate local file, which is theoretically the most robust approach, and layers on top a hooks system — essentially interception points set before and after every tool call (read file, run command, edit code) where execution can be forced to allow, deny, or prompt. It is the only one of the three that lets users impose hard constraints on AI behavior without modifying source code.
Industry View
Supporters argue that the hooks mechanism and external state files point in the right direction. Model capability will keep improving, but the real moat is "engineering encapsulation." Whoever first cracks the "AI runs for two hours without errors" experience breakpoint will capture the enterprise developer market.
The dissent is equally sharp. A senior engineer argues that this whole debate exposes a pseudo-need baked into the category — those slick two-hour "AI coding" demos mask countless manual rescue operations behind the curtain. Hooks may be powerful, but pushing complexity onto users amounts to making developers reinvent "operating system permission management" instead of solving actual productivity problems. Others caution that all three tools are iterating fast, and today's architectural differences may not exist in three months; betting on a specific product carries real risk.
Impact on Regular People
For enterprise IT: If you are already evaluating "let AI assist the dev team," the critical decision is no longer which model to pick — it is who can reliably complete a full requirement cycle, or the labor you save will be eaten by firefighting.
For individual professionals: Non-programmers need not worry yet, but "AI unreliable on long tasks" is a universal phenomenon. Any expectation of AI autonomously finishing an entire end-to-end job remains premature.
For the consumer market: The AI assistants consumers actually encounter (customer service bots, office assistants) suffer from the same "loses coherence after a long session" problem. The underlying mechanism is the same technical debt discussed here, which means expectation management matters more than feature checklists.