Work

Introducing Grounding into the Agentic Design Loop: An Exploratory Review of an Unclaimed Problem

Year: 2026
Type: preprint
Abstract: Large language model (LLM) based agents are moving from executing well-specified tasks to Agentic Design: shaping what and how to build upstream of fabrication, integration, and deployment. However, an internally coherent but externally infeasible design propagates its compounding error downstream. The failure is easy to overlook, yet consequential and hard throughout: a design-generating loop rarely includes a check against the world, and the error surfaces only after the artifact is built. Design Grounding names the correspondence between such a design and the world it must satisfy. The mechanisms that establish it exist in adjacent fields but have been introduced into the design loop only in isolated, domain-specific forms, which leaves the same difficulty scattered across fields under different names. Studying both sides through their own corpora, the review is the first to consolidate them into one unclaimed problem. What makes the problem hard is structural: a design's referent is a system that has not been built, and the model that produces the design cannot supply it without external intervention. The review is exploratory in aim: it characterizes the problem and screens Agentic Design and Grounding in all four combinations. Reading titles and abstracts with an LLM flags roughly one in fifteen of nearly three thousand papers as closing the Design Grounding loop; only twenty survive full-text adjudication and manual review. Those twenty sort into four structural difficulties, and each becomes an open problem whose parts already exist in an adjacent field but remain unassembled.
Institution: Tongji University
Language: English
Keyword: LLM agents
Open Access status: green
OpenAlex ID: W7225886379
Cites: 74
Cited by:
Related to: 10