8小时。0.12美元。
8月14日，Writer发布了Palmyra X6企业模型。两个数字最值得关注：智能体可以围绕单一目标无人值守运行8小时，平均每项任务成本约0.12美元。
第一个数字解决的是Agent能不能长时间干活的问题，第二个数字解决的是Agent用不用得起的问题。两个问题同时解决，长时Agent的商业化门槛，就被大大降低了。
这听起来可能不性感——不就是跑8小时吗？但真正做过Agent的人都知道，长时间运行不漂移，比短时间答对难题难多了。
最难的不是上下文长，是不忘初衷
很多人以为，长时Agent的瓶颈是上下文窗口。只要上下文够长，Agent就能跑很久。
大错特错。
真正的问题是意图漂移——Agent跑着跑着，就忘了自己最开始要干嘛。它可能在第一步、第二步、第三步都做得很好，但到了第十步、第二十步的时候，目标就歪了。它可能开始纠结一些无关紧要的细节，可能走进了死胡同还在死磕，甚至可能完全偏离了最初的任务目标。
人也会这样。你坐下来想写一篇文章，写着写着去查资料，查着查着刷起了新闻，刷着刷着看起了视频——三个小时过去了，文章一个字没写。这就是意图漂移。
人有自我意识，可以把自己拉回来。Agent没有。一旦跑歪了，它只会越跑越远，直到把token烧完为止。
"Agent最大的挑战从来不是上下文不够长，而是跑着跑着就忘了自己要干嘛。8小时不漂移，才是真正的长时Agent。"—— 一位AI Agent开发者
Palmyra X6的核心突破就在这里：它能在8小时的长任务中，保持目标的一致性。不会做着做着跑偏了，不会在无关的事情上浪费时间和token。它会规划、执行、测试自己的输出、修正、交付——整个过程不需要人盯着。
这意味着什么？意味着Agent可以真正进入后台自动化的场景了。你不用守在屏幕前看着它干活，可以把它扔在后台跑一宿，第二天早上来看结果。
比如，监控竞品动态、处理批量数据、整理行业信息、生成周报月报……这些以前需要人花好几个小时做的事情，现在可以交给Agent在后台自动跑。
一毛二一个任务的经济学
再来说成本。0.12美元一个任务——换算成人民币，大概八毛多钱。
这个价格是什么概念？如果你让一个实习生做同样的工作，一小时至少要十几二十块钱。而Agent做一个任务只要八毛多，而且可以24小时不间断地做。
更关键的是，Palmyra X6是基于GLM-5.2做的后训练——也就是说，它用的不是最顶级的基座模型，而是通过后训练和工程优化，在Agent场景下做到了极高的性价比。
这指向了一个重要的行业趋势：企业级AI不需要最强的通用模型，需要的是最合适的场景模型。通用大模型什么都能做，但什么都贵。而针对特定场景优化过的模型，可能能力没那么全面，但在它擅长的领域，效果差不多，成本却低很多。
对于企业来说，这才是真正的AI落地路径。不是买一个最贵的大模型什么都用它，而是根据不同的场景，选择不同性价比的模型。该用旗舰的地方用旗舰，该用专用模型的地方用专用模型——组合优化，才是成本最优解。
当一个Agent任务只要几毛钱、能在后台跑8小时的时候，企业就有动力把越来越多的重复性工作自动化。而每多一个自动化的场景，就多一个数据回流的闭环，模型就会变得更强——这是一个正向循环。
长时Agent的时代，正在从概念走向落地。
明天见。
8 hours. $0.12.
On August 14, Writer launched the Palmyra X6 enterprise model. Two numbers matter most: agents can run unattended toward a single goal for 8 hours, and the average cost per task is about $0.12.
The first number solves whether agents can work for extended periods. The second solves whether agents are affordable. Solve both at once, and the commercialization threshold for long-running agents drops dramatically.
This might not sound sexy - so what if it runs 8 hours? But anyone who's actually built agents knows: running long without drifting is way harder than acing short tasks.
The Hardest Part Isn't Context Length - It's Staying On Target
Many people think the bottleneck for long-running agents is the context window. Just make context long enough, and agents can run forever.
Dead wrong.
The real problem is intent drift - as an agent runs, it forgets what it was originally supposed to do. It might do great on step one, step two, step three - but by step ten, step twenty, the goal has wandered. It might start obsessing over irrelevant details, walk into a dead end and keep pushing, or completely diverge from the original task objective.
Humans do this too. You sit down to write an article, then you go look something up, then you're scrolling the news, then you're watching videos - three hours later, you haven't written a word. That's intent drift.
Humans have self-awareness and can pull themselves back. Agents don't. Once they go off track, they just keep going further until they burn through all their tokens.
"The biggest challenge for agents was never short context windows - it was forgetting what they were doing mid-task. 8 hours without drifting is what a real long-running agent looks like." - An AI Agent Developer
Palmyra X6's core breakthrough is exactly this: it can maintain goal consistency across 8-hour long tasks. It doesn't drift off mid-task, doesn't waste time and tokens on irrelevant things. It plans, executes, tests its own output, corrects, and delivers - the whole process without human supervision.
What does this mean? It means agents can truly enter background automation scenarios. You don't have to sit in front of the screen watching it work - you can leave it running in the background overnight and check the results in the morning.
For example: monitoring competitor activity, processing batch data, compiling industry information, generating weekly or monthly reports... things that used to take humans hours can now be handed off to agents running automatically in the background.
The Economics of 12 Cents Per Task
Now about cost. $0.12 per task - in most currencies, that's pocket change.
What does that price mean? If you had an intern do the same work, it'd cost you at least a few dollars an hour. An agent does a task for 12 cents, and it can do it 24/7 without breaks.
More importantly, Palmyra X6 is based on post-training on GLM-5.2 - meaning it's not using the most premium base model. It's achieving extremely high cost-performance in agent scenarios through post-training and engineering optimization.
This points to an important industry trend: enterprise AI doesn't need the strongest general model - it needs the most suitable model for each scenario. General foundation models can do everything, but everything is expensive. Models optimized for specific scenarios might be less versatile, but in their domain the results are similar and the cost is much lower.
For enterprises, this is the real AI落地 path. Not buying the most expensive foundation model and using it for everything, but choosing different cost-performance models for different scenarios. Use flagships where you need flagships, use specialized models where specialized models work - combinatorial optimization is the cost-optimal solution.
When an agent task costs pocket change and can run 8 hours in the background, enterprises have incentive to automate more and more repetitive work. And every additional automated scenario adds another data feedback loop, making the model stronger - it's a virtuous cycle.
The era of long-running agents is moving from concept to reality.
See you tomorrow.
Agent最大的挑战从来不是上下文不够长，而是跑着跑着就忘了自己要干嘛。8小时不漂移，才是真正的长时Agent。
—— 一位AI Agent开发者
The biggest challenge for agents was never short context windows - it was forgetting what they were doing mid-task. 8 hours without drifting is what a real long-running agent looks like.
- An AI Agent Developer
