From Context to Code: Codifying Cross-System Reasoning for LLM-Assisted Operations

Abstract

Large language model (LLM)-assisted operations often rely on many independent tools, with the operator or agent stitching partial answers together client-side. This can produce correct answers, but the reasoning remains brittle, session-local, and difficult to reuse. We present Axiom, a production federated query engine at Atlan, as one industrial instantiation of a broader pattern: codifying recurring operational knowledge as executable functions over a federated, composable query interface, so that humans and agents search, invoke, and compose shared operational definitions rather than rediscovering multi-step logic. The pattern has three ingredients: a federated, composable executable interface across heterogeneous sources; a curated library of operational definitions; and a search-before-write discipline that steers consumers toward reuse. We ground it in three production examples (incident diagnosis, fleet-wide targeting, and rollout eligibility) and in a controlled evaluation across nine models and two agent harnesses. The Axiom (federated-executable) condition passed 88.1% of cells versus 50.4% for the native-tool baseline, and per-model pass rates narrowed from a 27–73% baseline range to 73–93%. A stratified query ramp locates the effect: the largest gap (11.1% to 70.4%) occurs on a three-source query for which no library function existed, indicating that the advantage is driven principally by the federated executable substrate rather than by reuse of committed functions. The accumulation of a shared library into organizational memory is a production observation that this experiment does not isolate. We also discuss deployment lessons and the main open problem: versioning and evolution of shared operational definitions.