Closed-Loop Open-Ended AI Systems Are the Future: Meet DOLPHIN

Closed-loop open-feedback systems are the future.

Having worked on several AI projects, I’ve found that most approaches stop at idea generation, missing the validation step that makes real progress possible.

One thing we often overlook in agentic system development is the importance of feedback mechanisms.

In software engineering terms, think of it like a continuous integration/continuous deployment (CI/CD) pipeline: if you only integrate (idea generation) but never deploy and monitor in production (experimental validation), you can’t meaningfully improve your system.

Effective next-generation AI needs continuous feedback loops that validate ideas in near real-time, refine them based on actual performance metrics, and iterate rapidly.

For example, if you have a look at the DOLPHIN paper, you can see that the quality of generated ideas improves significantly when the system learns from experimental results.

The improvement rate jumps from 28.5% in the first loop to 50% by the third loop — a pattern with major implications for enterprise-scale AI solutions.

When building the system, researchers focused on three core components:

(1) idea generation,

(2) experimental verification,

(2) results feedback