From Stateless Code Generation to Living Ontologies Business Autonomous Entities in Action

Abstract

Most LLM-based engineering pipelines remain stateless: when requirements change, they regenerate code afresh and discard the vocabulary that ties schemas and APIs to business meaning. We address this limitation with Business Autonomous Entities (BAEs), ontology-aware agents that embed a living ontology of a business concept and orchestrate domain-preserving evolution. We evaluate BAEs with a two-tier design. A preliminary feasibility demonstration shows that a BAE can perform a six-step evolution online, without restarts. A fully automated benchmark then compares BAEs to GHSpec and ChatDev under identical prompts, seeds, and model configuration (gpt-4o-mini). For each framework, an identical six-sprint program is executed 100 times (serially), with programmatic clarification and post-hoc reconciliation of usage data. In addition, we apply a lightweight effectiveness sanity check that verifies whether each run produces a runnable project satisfying explicit structural requirements from the prompt. We analyze end-to-end execution time, token demand, monetary cost computed from reconciled tokens (including cached-input discounts), and LLM interaction counts. Across 300 runs, BAEs achieve substantially lower time, token usage, cost, and interaction counts than both baselines; effect sizes are predominantly medium to large, and BAEs occupy a Pareto-favorable region in the time–cost plane. A regression of total tokens on total cost, in combination with the cached-input pricing model, explains cross-framework cost differences via token volume and cache utilization. These results establish ontology-aware orchestration as a practical route to efficient and reproducible evolution under explicit structural constraints—shifting LLM-based development from one-shot generation toward continuous, domain-aligned change.