Linghao Zhang张凌豪
中文
Back to publications
1/2 The integrated Nurture-First Development framework.
2/2 The three-layer cognitive architecture.

arXiv2026PreprintSole author

Nurture-First Agent Development: Building Domain-Expert AI Agents Through Conversational Knowledge Crystallization

Linghao Zhang

arXiv preprint, 2026

arXivPDF

Abstract

The emergence of large language model (LLM)-based agent frameworks has shifted the primary challenge in building domain-expert AI agents from raw capability to effective encoding of domain expertise. Two dominant paradigms—code-first development, which embeds expertise in deterministic pipelines, and prompt-first development, which captures expertise in static system prompts—both treat agent construction as a discrete engineering phase preceding deployment. We argue that this sequential assumption creates a fundamental mismatch with the nature of domain expertise, which is substantially tacit, deeply personal, and continuously evolving. We propose Nurture-First Development (NFD), a paradigm in which agents are initialized with minimal scaffolding and progressively grown through structured conversational interaction with domain practitioners. The central mechanism is the Knowledge Crystallization Cycle, whereby fragmented knowledge embedded in operational dialogue is periodically consolidated into structured, reusable knowledge assets. We formalize NFD through: (1) a Three-Layer Cognitive Architecture organizing agent knowledge by volatility and personalization degree; (2) the Knowledge Crystallization Cycle with formal definitions of crystallization operations and efficiency metrics; and (3) an operational framework comprising a Dual-Workspace Pattern and Spiral Development Model. We discuss the conditions, limitations, and broader implications of NFD for human-agent co-evolution.

Cite

Linghao Zhang, “Nurture-First Agent Development: Building Domain-Expert AI Agents Through Conversational Knowledge Crystallization”, arXiv preprint, 2026.

Deep dive

Don't build an expert agent. Raise one.

A development paradigm in which an agent starts with minimal scaffolding and grows through structured conversation with the person who uses it.

What breaks

Most domain agents are built in one of two ways. Code-first development embeds expertise in deterministic pipelines. Prompt-first development captures it in a long system prompt. Both treat construction as a phase that ends before deployment.

That assumption does not fit how expertise works. It is substantially tacit, personal and still changing, so it resists being written down in advance.

The idea

Make development and deployment concurrent. The agent begins small and grows through daily use. The central mechanism is the Knowledge Crystallization Cycle: fragments of knowledge in everyday dialogue are periodically consolidated into structured, reusable assets, which then shape the next round of conversations.

Plate 1 The framework: a three-layer knowledge architecture inside a four-phase crystallization cycle, with two workspaces and a spiral over time.

The architecture

Knowledge is organized by how often it changes and how personal it is. A Constitutional layer holds identity and rules, and loads in every session. A Skill layer holds modules and references, and loads on demand. An Experiential layer holds logs and cases, and is searched semantically.

Plate 2 The three layers, with crystallization moving knowledge upward and grounding moving it back down.

Two operating patterns follow. A Dual-Workspace Pattern separates a surgical workspace for batch changes from a nurturing workspace for daily conversation, over one shared file system. A Spiral Development Model alternates scaffolding, nurturing and crystallization.

Where it fits

The paper does not claim the paradigm wins everywhere. It fits when expertise is largely tacit, differs from one practitioner to the next and keeps changing, when the user works with the agent through dialogue, and when recalling the user’s own past decisions adds value. Where expertise can be fully formalized and does not change, as in tax-form processing, code-first development is the better choice.

What is left

Early on, the agent offers little, which can discourage the sustained engagement it needs to become useful. The quality of nurtured knowledge is hard to assess objectively, and the agent may absorb its user’s biases along with real expertise. Crystallization still needs semi-manual direction. Establishing the paradigm’s advantage in numbers needs controlled studies with several practitioners, common evaluation criteria and code-first and prompt-first baselines. One of my open repositories sits close to this idea: personal-knowledge-wiki.

100%

↑↓ Move↵ Open