Nurture‑First Agent Development: arXiv paper argues for conversational crystallization to build domain‑expert AI

Lead: a shift from code and prompts to conversation

A new arXiv preprint, arXiv:2603.10808, argues that the hard part of building domain‑expert AI agents is no longer raw model capability but the effective encoding of specialized knowledge. The paper contrasts two current development paradigms — code‑first systems that hard‑wire expertise into deterministic pipelines, and prompt‑first systems that try to capture expertise inside heuristics and prompt engineering — and proposes a third path: "nurture‑first" development, driven by conversational knowledge crystallization. The claim: let dialogue itself distill, structure and modularize expertise so agents become easier to maintain and transfer across domains.

What the method promises

At its heart the approach treats conversations between humans and models as a medium for iterative refinement. Rather than burying domain logic in brittle code or opaque prompt stacks, knowledge crystallization surfaces assumptions, edge cases and rules through back‑and‑forth interaction, producing artifacts (structured Q&A, canonical examples, tool interfaces) that can be consumed by agent frameworks. The paper positions this as especially useful for regulated or high‑stakes fields such as healthcare, finance and legal work, where auditability and provenance matter as much as raw performance. Note that this research is a preprint and has not been peer reviewed.

Why this matters now — and for China’s tech players

Why should Western readers care? Agent frameworks are rapidly being commercialized worldwide, and encoding domain expertise cleanly is a bottleneck for deployment at scale. It has been reported that major Chinese players are aggressively pursuing agent strategies: Baidu (百度), Alibaba (阿里巴巴) and Huawei (华为) have all invested in LLM platforms and agent toolchains, reportedly tailoring them for domestic industry partners and regulated markets. Geopolitics also colors the conversation: export controls and trade policy are incentivizing localized capability and robust, auditable knowledge representations that can be deployed under restrictions.

Next steps and caveats

The paper presents a conceptual and empirical case for nurturing agents via conversational crystallization, but questions remain about automation, labor costs, and how to validate distilled knowledge at scale. Will human-in-the-loop refinement scale for enterprise needs? Can crystallized artifacts be standardized across platforms? Those are practical hurdles before the idea moves from academic proposal to industry practice. For now, the preprint offers a fresh vocabulary and a design pattern that could reshape how organizations turn LLMs into trustworthy, domain‑expert agents.