GABBE: A Neurocognitive Swarm Architecture for Agentic AI Software Engineering

The transition from human-centric to agent-centric software engineering has revealed critical limitations in existing Multi-Agent Systems (MAS), particularly concerning context decay, non-determinism, and cost asymmetry. This paper introduces the Generative Architectural Brain Base Engine (GABBE), a third-generation autonomous software system built upon neurocognitive principles. GABBE features a dual-layer architecture: a meta-cognitive "Brain Mode" utilizing Active Inference to minimize project risk, and a "Loki Swarm Mode" executing deterministic, spec-driven development life cycles across 30+ specialized agent personas. This paper details the system's 4-layer memory hierarchy, cost-effective Large Language Model (LLM) routing, and the programmatic enforcement of the "Another Lethal Trifecta" mitigation strategy. A comparative analysis demonstrates GABBE's architectural superiority over flat-topology orchestrators, providing a scalable blueprint for applying neurocognitive agentic patterns to zero-to-one software development.