# Neuro-Symbolic Inversion of Control

> Until roughly 2023, many efforts used LLMs by calling OpenAI-compatible APIs from conventional programs. In 2025, the approach known as “AI agents and tools,” in which LLMs call conventional programs, is gaining popularity. Some even call 2025 “the year of agents.”

Until roughly <2023>, many efforts used [LLMs](https://wiki.g15e.com/en/pages/Large%20language%20model.txt) by calling <OpenAI>-compatible <APIs> from conventional programs. In <2025>, the approach known as "AI agents and tools," in which [LLMs](https://wiki.g15e.com/en/pages/Large%20language%20model.txt) call conventional programs, is gaining popularity. Some even call 2025 "the year of agents."

## Explanation

```mermaid
sequenceDiagram
	autonumber
	actor U as User
	participant C as Code
	participant L as LM
	U ->>+ C: 실행
	loop Code의 로직에 의해 반복
		C ->>+ L: 호출
		L -->>- C: 생성된 출력
	end
	C -->>- U: 결과
```

If the conventional approach above represents forward control, the design below, where an AI agent uses tools, can be viewed as a kind of <inversion of control>.

```mermaid
sequenceDiagram
	autonumber
	actor U as User
	participant C as Code
	participant A as Agent
	
	U ->>+ A: 명령
	loop Agent가 끝났다고 판단할 때까지
		A ->>+ C: 호출
		C -->>- A: 결과
	end
	A -->>- U: 생성된 출력
```

The AI agent controls the high-level flow, while tools perform specific low-level tasks when called by the agent.

Since conventional code (symbolic computation) and a [LLM](https://wiki.g15e.com/en/pages/Large%20language%20model.txt) (a neural network) collaborate, this belongs to neuro-symbolic computation. In particular, the latter arrangement reverses control, so we might call it the **"Neuro-Symbolic Inversion of Control" pattern**.

## References

- [Recursive Neuro-Symbolic Inversion of Control](https://wiki.g15e.com/en/pages/Recursive%20Neuro-Symbolic%20Inversion%20of%20Control.txt)