What This Is
This week, a hands-on post in the Reddit tech community sparked discussion: the author let Strata (an AI coding tool based on the Qwen 3.8 model) run automated coding continuously for 12 hours, and found two typical failure modes in the second half — "memory fragmentation" and "tool hallucination" (AI fabricating tool return results out of thin air). This exposed the reliability shortcomings of current AI Agents (AI capable of autonomously executing multi-step tasks) on long tasks.
Specifically: the first type, "memory fragmentation," is when AI finishes writing an execution report and immediately asks itself "Did I really write this report?"; the second, more dangerous type — AI reads a file path and misses one digit, repeatedly tries this wrong path, and then blames the tool, saying "you returned a corrupted path."
Industry View
The tech community generally acknowledges this is a common ailment of current AI Agents. OpenAI, Google, and Anthropic have all been investing in "context compression," "long-term memory," and "self-verification" over the past six months, but none has solved it at the root. Supporters argue that Strata performs adequately on short tasks (1–2 hours); the problem is concentrated in "continuous ultra-long tasks" and is an engineering issue, not a principle issue — a few iterations could fix it.
The opposing voice is sharper: a senior developer in the comments noted that this precisely shows current Agents are fundamentally unsuitable for "unattended overnight tasks." If enterprises integrate them into production pipelines, they must retain a manual spot-check step — otherwise a single hallucination could delete the wrong files or deploy the wrong version. We lean toward the latter view: a 0.1% reliability flaw in production environments will be magnified into a 100% accident.
Impact on Regular People
For enterprise IT: when evaluating AI coding tools, "how many hours can it write continuously without errors" matters more than "can it write at all" — the procurement checklist needs a new "long-task stability" metric.
For individual professionals: treating AI as a coding assistant is fine, but treating it as an "independent coworker who delivers finished work" is premature — you can't let it run overnight and submit its results directly the next day.
For the consumer market: AI coding product marketing over the next year or two will bifurcate — one type will emphasize "speed" (short-task efficiency), another will emphasize "reliability" (long-task dependability), and the latter may be priced significantly higher.