LLMoxie platform turns AI coding agents into domain-aware science collaborators

LLMoxie, presented at a SIGKDD 2026 workshop, is an institutional AI platform designed to bridge the gap between off-the-shelf AI coding agents and the needs of scientific software development. Developed by a university-based research software engineering (RSE) center over 20 months, it directly addresses the chronic misalignment of commercial agents that prioritize raw code quality metrics but neglect scientific conventions—such as proper library usage, data sensitivity and embargo handling, and producing decision trails that can be audited, cited, and reproduced. The platform’s three-tiered architecture includes a multi-cloud/on-premise inference layer, a control plane (based on LiteLLM and MLflow) for authentication, budgeting, PII masking, and observability, and an application augmentation layer for AI coding agents.

Layered on top is an open-source RSE-Plugins ecosystem that encodes accumulated research software engineering knowledge as a Plugin-Agent-Skill hierarchy. This hierarchy spans scientific Python best practices, domain-specific knowledge (from astronomy to health), a six-phase research-and-implement workflow, and project lifecycle management. The result is that LLMoxie transforms generic code generators into domain-aware collaborators that respect community norms and produce technical reasoning with auditable provenance. The paper reports operational lessons from real deployments across astronomy, earth and climate science, agriculture, and health projects, highlighting recurring challenges in infrastructure, governance, and process for adopting agentic AI in multi-domain RSE centers.