AI Agents: Give 'Goals' Only, No 'Interference'

The latest agentic models invalidate past methods of detailed prompt writing. Because models excel at self-reasoning and planning, overly specific instructions or examples can actually degrade their performance. It is now crucial to clearly define the model's goals and success criteria quantitatively. If users do not precisely indicate 'when inference should stop,' the model may continue unnecessary reasoning. Therefore, it is important to actively utilize options to control the model's inference effort.

The current AI market is restructuring, moving away from dependence on specific models to using various models segmented by business purpose. In the past, many companies relied on a single large model, but now the trend is to select the most suitable model for the nature of the task. For example, GPT-based models are preferred for tasks like document classification or summarization, while models like Claude are favored for creative writing or story generation. This shift maximizes the strengths of each model and contributes to overall business efficiency.

Latest AI models like Fable 5 tend to be excessively proactive, performing tasks unnecessarily even with slightly ambiguous user queries. This excessive proactiveness leads to 'token burning,' causing unnecessary cost consumption. For instance, a vague query can frequently result in the model exploring more data than needed and validating various hypotheses, consuming an enormous number of tokens. Therefore, a key challenge in prompt engineering has become appropriately controlling this model proactiveness and setting clear boundaries to prevent unnecessary token consumption. The ability to control the model through user instructions to prevent it from over-extending itself has become crucial.

Zhipu AI's GLM 5.2 model stands out for its long-duration task performance. In the developer community, there are reports of the model running continuously for over 10 days to perform long-term coding tasks, showcasing its excellent sustained performance. Thanks to this characteristic, GLM 5.2 is highly regarded as particularly useful for complex and continuous coding projects, gaining significant popularity among developers. This illustrates the emergence of specialized AI models optimized for specific purposes.

One of the recent trends in AI model utilization is 'token maxing,' a method of using a large volume of tokens to achieve sophisticated outputs. Expressions like 'Luna maxing' and 'Kimi maxing' are gaining popularity, indicating that certain models are designed to allocate a large number of tokens for deeper reasoning. Especially, the latest models, which tend to use more tokens compared to US models, are employing this 'maxing' approach to provide higher quality results to users. This phenomenon can be seen as a result of both increased model processing capabilities and higher user expectations.

When using AI models trained on Chinese data, it's common for Chinese characters or phrases to appear in responses, even if the query is in Korean. This is because a significant portion of the model's training data includes Chinese. To resolve this, users can adjust the model's temperature setting or add output constraints. Specifically, efforts are needed to reduce language mixing by explicitly limiting non-Korean language output in prompt settings.

There is a growing trend of tightening regulations on AI data center construction in various regions, including Texas. This measure is interpreted as an effort to address the immense power consumption problem associated with AI model operations. As physical resource and power supply limitations for operating high-performance large models become clear, it may become difficult to insist on only the highest-performing models. Consequently, demand for lighter models that consume less power and can run efficiently while delivering similar performance is expected to increase in the future AI market. This will likely also impact the direction of AI technology development.

As inference models have highly advanced, detailed instructions and step-by-step thinking guidance, such as 'think step by step,' which were used in the past, are no longer effective. Inference models have already entered the 'zero-click' era, possessing the ability to plan and execute tasks autonomously once the user's intent is clearly conveyed. Therefore, the focus should now be on clearly communicating the situational context and the ultimate goal to the model. Old-fashioned prompts will inevitably fall short in both efficiency and accuracy, so prompt writing methods must shift to align with the changed characteristics of the models.

In the past, it was important to instruct AI models to think more, reason more, and act in detail, but this approach has changed. For the latest inference models, it is more efficient to clearly present the exact goal, current situation, and context of the task rather than providing short and concise instructions. The key is to provide this information so the model accurately understands what it needs to do. Unnecessarily detailed instructions can actually degrade the model's autonomous reasoning capabilities.

Because inference models have the characteristic of autonomously forming their reasoning plans, they can continue unnecessary tasks indefinitely if the user does not set clear termination conditions. If a model starts with 'I will infer,' the user must specifically designate until when this inference should be performed. Otherwise, the model may continue the task, causing unnecessary resource consumption. Therefore, for efficient use of the model, it is essential to specify not only the ultimate goal of the task but also conditions for termination upon completion or achievement of a specific criterion.

Recently, various options have emerged to adjust the intensity of inference, or the 'effort' of inference, in AI models. This value should be understood not merely as limiting the amount of tokens used, but as controlling how deeply the model will deliberate and infer a specific task. For example, for simple tasks like classification or information extraction, which are not complex, inference intensity can be set low to increase efficiency. Conversely, for complex problem-solving or long-term project planning tasks that require deep thought, inference intensity can be set high to encourage the model to spend sufficient time deriving optimal results. This adjustment is crucial for balancing resource consumption and output quality.

A new effective way to utilize inference models is to switch from 'requesting review' of the output to 'requesting evidence.' Previously, it was common to ask the AI to review its results once more, but with the latest inference models already undergoing several internal review processes to derive the final output, such re-review requests are inefficient. The model has already done its best to review. Instead, it is more important to ask the model to explain the process or evidence that led to a specific result. This allows the user to understand the model's reasoning process, detect potential errors, or enhance the reliability of the results. This is a key method for improving the model's utility.

It has been argued that providing detailed examples to an inference model can actually be detrimental. This is due to 'example bias,' where the model gets stuck within the provided examples and tends to remain within their scope, failing to produce more creative or autonomous responses. One of the biggest advantages of inference models is their ability to think autonomously and explore various possibilities, but detailed examples hinder this advantage. Therefore, to encourage the model to infer broadly and find new solutions, it is important to avoid providing unnecessarily specific examples.

When inputting large volumes of data, such as 1 million tokens, into an AI model, it is essential to structurally and clearly separate the data from the instructions. Simply putting all content together can cause the model to confuse data and instructions, leading to important instructions being skipped or not properly recognized. To prevent this, you must clearly mark the beginning and end, such as 'data starts here and ends here.' Using XML tags or other delimiters to visually distinguish data sections from instruction sections, and placing the core instruction to be performed on the last line, is an effective method. Such structured input helps the model accurately understand and process information.

The latest agentic models can perform long-term tasks autonomously, but there is also a possibility of unexpected errors occurring during this process. Therefore, instead of the past prompt method focused on detailed instructions, users should instruct, 'Before you touch anything, first tell me your plan of what you are going to do.' It is more efficient to have the model first report its work plan to the user, and then the user reviews and approves that plan. This shift respects the model's autonomy while allowing the user to maintain ultimate control, preventing unnecessary errors or resource waste.

There's no need to be obsessive about using specific delimiters like XML tags when inputting information into AI models. The important thing is that any symbol clearly separates information and instructions for the model. Korean notation or other symbols can be used to distinguish data from instructions, or to separate various sections. The core is to help the model accurately recognize each context without confusing information. Therefore, users should use delimiters in the way they find most clear. This flexibility enhances the convenience of prompt writing and improves communication efficiency with the model.