Summary of Key Points
This news article focuses on the engineering evolution of AI Agents, which are capable of completing complex tasks autonomously. Zhang Yutao, co-founder of Yuezhi Dianmian (Moon's Dark Side), shared three stages of technological advancements: Prompt Engineering in 2023, Context Engineering in 2024, and Harness Engineering in 2026. He also discussed the "scaffold philosophy" – which involves removing auxiliary tools as the model's capabilities improve – along with four core principles and the key metric for evaluating an Agent's value: human trust.
The Three Stages of Agent Evolution: From "Asking Questions" to "Doing Things autonomously"
The enhancement of AI Agents is driven by three iterations in engineering methods:
- 2023: Prompt Stage (4K Context – A ‘Small Note’)
At that time, ChatGPT had only a context capacity of 4K, which was akin to the size of a small note, allowing it to perform basic question-and-answer tasks. To make it more capable, questions needed to be crafted precisely. For example, asking "Write an article about AI" was insufficient; instead, one would need to specify "Write a 300-word AI科普 article for elementary school students, using examples of robots." The essence was to provide precise instructions within a limited context to match the AI's understanding capabilities.
- 2024: Context Stage (Expanded Context + Tool Kit)
The model's context window expanded significantly (from 4K to 128K or more), enabling it to handle more complex information. This led to the introduction of "tool invocation" – for instance, asking the AI to check the weather or calculate data. The instructions for using these tools were included in the context, allowing the AI to decide when to apply them. For example, if you asked the AI to "plan a weekend trip to Beijing," it would automatically use map and weather apps to gather information and provide a plan.
- 2026: Harness Stage (Circular Laboratory)
AI is now capable of handling hour-long tasks, such as writing a comprehensive market report. However, this requires a "circular environment" where the AI attempts the task, identifies issues (e.g., insufficient data), decides on the next steps (e.g., querying industry databases), executes the task, and adjusts its approach based on the results. This process is like providing an AI with a laboratory where it can iterate and optimize autonomously, without relying on a single instruction.
The Essence of an Agent: A "Loop" + The "Scaffold Philosophy"
Although AI Agents may seem complex, their core functionality revolves around a "while loop" – a common programming concept for repetitive execution:
- Loop Logic: If the task is not completed → Consider the next step → Execute → Check the results → Repeat the process until the task is finished. For example, when writing a report, if the AI realizes it lacks competitor data, it will search for it and continue with the report until it is complete.
- Scaffold Philosophy: This is similar to how children learn to walk with assistance. When the model's capabilities are limited, auxiliary tools (e.g., smaller models for preliminary filtering) are used to help it perform tasks. As the model improves, these tools are removed to prevent them from becoming a bottleneck.
The Four Core Principles of Agent Engineering
Zhang Yutao’s team has identified four key principles based on practical experience:
- Principle 1: Keep the Agent Simple
The essence of an AI Agent is its loop logic; adding overly complex designs can hinder its performance. Simplicity allows the model to function more effectively.
- Principle 2: Include Relevant Information in the Context
The context should contain only essential information, such as task objectives and tool instructions. For example, when asking the AI to write a report, provide it with relevant industry data rather than outdated news.
- Principle 3: Clear and Unambiguous Tool Design
Tools provided to the AI (e.g., data retrieval interfaces) should be as user-friendly as smartphone app buttons – with clear functions and step-by-step instructions. Ambiguous instructions can lead to misuse or abandonment by the AI.
- Principle 4: Balance Memory Between Consistency and Creativity
AI has limited memory; therefore, it’s important to balance the need to remember previous steps (consistency) with the ability to innovate flexibly. For instance, important intermediate results can be summarized and included in the context to save space while still conveying essential information.
The Key to Evaluating Agent Value: Human Trust
The true value of an AI Agent is determined by whether humans are willing to entrust it with complex tasks. For example, would you trust an AI to organize a week’s worth of meeting minutes and generate an action plan on its own? If the AI can solve problems autonomously without errors within a loop-based process, it gains more trust and thus higher value.
The core idea is that the goal of an AI Agent is not to execute instructions perfectly once but to grow autonomously within a closed-loop system, gradually earning human confidence to handle increasingly complex tasks.
Conclusion
The evolution of AI Agents represents a shift from passive response to proactive action. The three iterations in engineering methods aim to enable AI to perform more freely. The ultimate measure of its success is the level of trust humans have in it – the greater the trust, the better the Agent can integrate into our work and daily lives.