What is Loop? It's been quite popular in recent days~

Loop is the agent loop: the model no longer ends with "receiving an instruction → outputting a result." Instead, it enters a loop—reason → act → observe results → reason again—until the task is truly completed before exiting. The structure on the right side of the diagram, inside the dashed box, that loops back is the loop.

Essentially, it brings a well-established practice in software engineering into AI: REPL, event loops, while loops—all structures that "execute repeatedly until a condition is met." Loop simply replaces the "loop body" with "model thinking + tool invocation."

Why can it replace prompts?

The key is not the word "loop," but rather that it shifts the burden of complexity from "your text" to "system structure." Let me break it down into four levels:

1. From "describing how to do it" to "defining what to do"
A prompt is a static encoding—you must pre-plan every step of a task and write it down in natural language. The problem is that for real-world tasks, you simply can't pre-plan everything. In an investment analysis, should you check financial reports first or look at the industry first? It depends on what you find. Prompts hardcode the path, while the loop lets the model decide the next step based on real-time feedback from each step. You only provide the goal and capability boundaries; the path is dynamically generated.

2. Context evolves from "being stuffed in" to "growing organically"
The power of a prompt depends on how much information you can cram into the context window—but you often don't know in advance whether that information is needed. In a loop, every tool call and every observation is automatically appended to the context. Context is not a one-time budget you plan; it grows dynamically with the loop. Whatever the model needs, it fetches.

3. Error correction capability is introduced
A single prompt has no concept of "retry"—wrong is wrong. The loop comes with feedback: if a tool reports an error, a search finds nothing, or numbers don't match, the model can detect and correct it in the next reasoning step. This is like upgrading from "handwritten assembly" to "a high-level language with breakpoint debugging"—the former must be written correctly in one go, the latter allows you to fix as you run.

4. "How to do it" is solidified in tools, while "what to do" is left to natural language
This is the most critical point. Traditional prompt engineering describes "which API to call, in what order, how to parse the response" using text, leading to massive ambiguity. In the loop paradigm, these are turned into tools—function signatures are precise, unambiguous interfaces. The model only needs to decide "when to call which tool," without guessing how to call it through natural language. The ambiguity of prompts is replaced by the determinism of tools.

The essence in one sentence
Prompt = You compress the entire execution logic into a block of text and bet it runs correctly in one go. Loop = You define the goal + provide a set of tools + let the model grow its own execution path through real interaction.

It doesn't "eliminate" prompts—system prompts (who you are, what you can use, what rules to follow) still exist and are important. What it eliminates is the need to "write a perfect instruction for every single task." Complexity shifts from "the length of a single prompt" to "the quality of system design + tool set," and the latter can be engineered, reused, and tested.

That's also why the focus of this wave of AI applications is shifting from "prompt engineering" to "agent/system design"—when you can solidify "how to do it" into tools and loop structures, natural language only needs to express "what to do," which is exactly what it's best at.