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

This technology memo describes CARE (Collaborative Agent Reasoning Engineering), a disciplined, stage-gated process for engineering large language model (LLM) agents that specifies behavior, grounding, tool orchestration, and verification through reusable artifacts rather than trial-and-error prompt iteration. CARE reframes agent building as an engineering discipline centered on explicit specifications. The process is defined as a three-party workflow among subject-matter experts (SMEs), developers, and required LLM-based helper agents, where helper agents facilitate conversion of informal domain intent into structured, reviewable artifacts that humans approve at stage gates.

CARE is motivated by the uneven performance landscape of LLMs, where results vary significantly based on a user’s ability to manage domain-specific constraints and verification steps, often widening the gap between novice and expert analysts. CARE targets this practical reliability gap that emerges from uneven LLM performance and user application. CARE operationalizes agent development by producing concrete artifacts: interaction requirements, domain-grounding specifications, tool interfaces and orchestration plans, reasoning policies and guardrails, prompt architecture, and evaluation criteria, so that agent behavior is specifiable, testable, and maintainable over time.

In a case study building a NASA Earth science data discovery agent that queries the NASA Common Metadata Repository (CMR) Application Program Interface (API), the CARE-designed agent outperformed a baseline with identical model and tool access in a two-gate evaluation achieving higher Recall@1 on a large synthetic benchmark (n=621; 71.7% vs 69.1%) and higher Recall@5 on an SME-created gold benchmark (n=43; 27.2% vs 20.2%). CARE shows measurable retrieval gains under both fast synthetic testing and SME-grounded verification. Thus, CARE is presented as a stage-gated, artifact-driven, three-party methodology involving helper agents that yields measurable performance improvements in a realistic, tool-driven scientific data discovery setting. CARE demonstrates a repeatable method with concrete evidence of improved outcome