We present Omni, a production multi-agent system that completes heterogeneous long-horizon tasks — tasks requiring coordinated live web browsing, Python execution, image generation, email dispatch, and file operations under a single user intent — using GPT-4.1-mini as its sole backbone model. The central architectural insight is Cortex: a deterministic finite-state machine orchestrator coordinating eleven specialized agents through four structural mechanisms: (1) a continuous plan-anchored correction loop; (2) a tiered HITL architecture with three semantically distinct intervention tiers; (3) a deterministic FSM speaker-selection function at zero token cost; and (4) memory-mediated capability chaining via OmniMemory. Production deployment across 2,170 sessions confirms: 81.3% plan re-invocation rate, 88.0% HITL trigger rate, 79.5% task execution rate. FSM: 0% routing error, 1.6 μs/transition. On 5 GAIA Level 2/3 tasks: 5/5 (100%). Planned next optimization: Agent-to-Agent (A2A) protocol for dynamic capability discovery.
Kumar et al. (Tue,) studied this question.