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

Linghao Zhang

cs.AIMar 11, 2026

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

Authors: Linghao Zhang

Organizations: Jiangsu Key Laboratory of Wireless Communications Nanjing University of Posts and Telecommunications (NJUPT) Nanjing, Jiangsu, China

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 illustrate the paradigm through a detailed case study on building a financial research agent for U.S. equity analysis and discuss the conditions, limitations, and broader implications of NFD for human-agent co-evolution.

Figures & tables

Figure 1: Integrated overview of the Nurture-First Development framework. The diagram unifies the four core contributions: ➀ the Three-Layer Cognitive Architecture (center stack) organizing knowledge by volatility—Constitutional (blue, low), Skill (teal, medium), and Experiential (amber, high); ➁ the Knowledge Crystallization Cycle (surrounding ring) with four phases—Conversational Immersion, Experiential Accumulation, Deliberate Crystallization, and Grounded Application; ➂ the Dual-Workspace Pattern (left: Surgical Workspace for crystallization; right: Nurturing Workspace for operational dialogue); and ➃ the Spiral Development Model (outer expanding trajectory) through which these components interact across successive development revolutions. Arrows trace the flow of knowledge: tacit expertise enters through conversation (right), accumulates as experiential records (bottom), is crystallized into structured skill references (left), and grounds future interactions at progressively higher baselines (top). Section numbers indicate where each component is formally developed.

Figure 2: Three paradigms of agent development. Code-First and Prompt-First follow a linear develop-then-deploy lifecycle with a hard boundary between construction and operation. Nurture-First dissolves this boundary: a spiral of scaffolding, nurturing, and crystallization phases interleaves development with deployment, enabling continuous knowledge growth throughout the agent’s operational lifetime.

Dimension	Code-First	Prompt-First	Nurture-First (NFD)
Knowledge encoding	Deterministic code	Static prompt text	Evolving memory files
Development phase	Before deployment	Before deployment	Continuous, concurrent
Expertise capture	Explicit logic only	Declarative snapshot	Tacit + explicit, evolving
Update mechanism	Code change + redeploy	Prompt editing	Conversation + crystallization
Personalization	Parameterization	Prompt customization	Deep experiential adaptation
Primary developer	Software engineer	Prompt engineer	Domain practitioner (user)

Table 1: Comparative analysis of agent development paradigms across ten dimensions.

Figure 3: Three-Layer Cognitive Architecture. Knowledge is organized into three layers by volatility and personalization degree. The Constitutional Layer (low volatility, loaded every session) contains identity and principles. The Skill Layer (medium volatility, loaded on demand) contains structured domain knowledge. The Experiential Layer (high volatility, searched semantically) contains accumulated operational experience. Upward crystallization arrows represent knowledge consolidation; downward grounding arrows represent interpretive application.

Figure 4: The Knowledge Crystallization Cycle. Four phases form an ascending spiral: (1) Conversational Immersion generates knowledge fragments through operational dialogue; (2) Experiential Accumulation logs and tags these fragments in persistent memory; (3) Deliberate Crystallization consolidates patterns into structured knowledge assets; (4) Grounded Application deploys crystallized knowledge in practice, generating new experiences at a higher baseline. Each revolution raises the agent’s knowledge fidelity.

Figure 5: The Dual-Workspace Pattern. The Surgical Workspace (left) provides full filesystem access for batch processing, crystallization, and skill development. The Nurturing Workspace (right) is the agent’s runtime environment for daily conversational interaction. Both operate on shared state—the agent’s file system—enabling seamless knowledge transfer between development and operational modes.

Figure 6: The Spiral Development Model. An expanding spiral alternates between four quadrants: Scaffold (surgical workspace builds structure), Nurture (conversational interaction accumulates experience), Crystallize (surgical workspace consolidates knowledge), and Nurture + (agent applies crystallized knowledge at a higher baseline). Inner revolutions correspond to early phases; outer revolutions to mature operation. Radial expansion represents growing knowledge depth.

Figure 7: NFD applied to financial research agent development. The timeline shows the spiral development trajectory: from initial scaffolding (market data skills, skeleton analytical framework) through nurturing phases (daily market discussions, earnings analysis) to crystallization checkpoints (extracting evaluation frameworks, risk heuristics). Below: progressive population of the three cognitive layers. Inset: growth in agent usefulness over 12 weeks.

Metric	Wk 1–3	Post-CC1	Wk 9–12	Post-CC2
Useful analyses (%)	38	52	71	74
Case recalls	2	5	12	15
Bias flags	0	1	4	5
Skill refs populated	2	4	6	8
Error patterns	6	8	10	12
Daily log entries	21	–	60+	–

Table 2: Agent development progression metrics.

Figure 8: Micro-process of conversational nurturing. Three representative interaction episodes illustrating how domain expertise is progressively encoded through the NFD cycle. Episode A : Semantic correction and framework refinement through error identification. Episode B : Strategy externalization through guided dialogue, where tacit investment logic is made explicit and crystallized into a reusable decision framework. Episode C : Emergent skill extension, where a new analytical capability arises organically from operational need. Arrows indicate knowledge flow between experiential logs, skill references, and constitutional principles.

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 705 papers

ReCreate: Reasoning and Creating Domain Agents Driven by Experience
Jan 16, 2026Zhezheng Hao, Hong Wang, Jian Luo +6LLM Agents LLM Agent Self-Improvement
Zhejiang University · State Key Laboratory of Blockchain and Data Security, Zhejiang University · Tencent +3
Collaborative Agent Reasoning Engineering (CARE): A Three-Party Design Methodology for Systematically Engineering AI Agents with Subject Matter Experts, Developers, and Helper Agents
Apr 30, 2026Rahul Ramachandran, Nidhi Jha, Muthukumaran RamasubramanianLLM Agents Human-AI Collaboration
Marshall Space Flight Center, Huntsville, Alabama · The University of Alabama in Huntsville, Huntsville, Alabama
KUPAS MASTER: Distilling the Tacit Expertise of Master Practitioners into Agent-Ready Experience Corpora
Sep 29, 2026Changmian Wang, Yuchao Ma, Xuchao Lu +13LLM Agent Skill Learning
Shanghai Kupas Technology Co., Ltd. and Tongji University

Jan 16, 2026J/K move · Enter open · S save

ReCreate: Reasoning and Creating Domain Agents Driven by Experience

Zhezheng Hao, Hong Wang, Jian Luo, Jianqing Zhang+5

Zhejiang University · State Key Laboratory of Blockchain and Data Security, Zhejiang University · Tencent+3

LLM Agents LLM Agent Self-Improvement Self-Evolving Agents LLM Agent Workflow Optimization

Open paper AskPDF arXiv

Large Language Model (LLM) agents are reshaping the industrial landscape. However, most practical agents still rely on human-designed scaffolds, making their creation and adaptation across diverse domains labor-intensive. This situation poses a central question: can we automatically create and evolve domain agents through interaction experience in real-world environments? While recent approaches have sought to automate agent creation, several treat scaffold optimization as a black-box procedure guided primarily by final performance metrics. Such strategies overlook critical evidence explaining why an agent succeeds or fails and often incur high computational costs. To address these limitations, we propose ReCreate, an experience-driven framework for creating and evolving specialized domain agents. ReCreate systematically leverages agent interaction histories to identify the causes of success and failure and guide targeted scaffold updates. Specifically, we introduce an agent-as-optimizer paradigm in which an optimizer agent actively inspects execution experience and iteratively refines the task agent's scaffold through three key components: (i) an experience storage and retrieval mechanism for on-demand evidence inspection; (ii) a reasoning-creating synergy pipeline that translates execution evidence into updates to prompts, workflows, tools, and memory; and (iii) hierarchical updates that consolidate instance-level refinements into reusable domain-level improvements. Experiments across thirteen evaluation settings spanning four domains show that ReCreate consistently outperforms human-designed agents and existing automated agent generation and self-evolving baselines, even when evolving from minimal seed scaffolds.

Abstract

LLM Agents

Explore similar work

ReCreate: Reasoning and Creating Domain Agents Driven by Experience

Collaborative Agent Reasoning Engineering (CARE): A Three-Party Design Methodology for Systematically Engineering AI Agents with Subject Matter Experts, Developers, and Helper Agents

KUPAS MASTER: Distilling the Tacit Expertise of Master Practitioners into Agent-Ready Experience Corpora