What question did this study set out to answer?

The aim is to resolve the failure modes of LLM-centric autonomous agents in critical environments by redefining their operational structure.

April 28, 2026Open Access

HYDRA-Ω: A Deterministic Multi-Agent Architecture with Typed LLM-Sensor, BDI-HTN Cognition, and Episodic Anti-Pattern Learning

Puntos clave

The aim is to resolve the failure modes of LLM-centric autonomous agents in critical environments by redefining their operational structure.
Developed HYDRA-Ω as a multi-agent architecture where LLM functions solely as a typed sensor.
Established a six-step BDI-HTN cognitive cycle with bounded interfaces and an offline-verified plan library.
Proved four formal properties and tested guarantees on a three-CVE benchmark.
Demonstrated that structural hallucination is impossible in the typed LLM-sensor design.
Showed effective performance of the BDI-HTN cognitive engine in plan selection under regulated conditions.
Achieved safety by construction through PDDL verification of the plan library.

Resumen

LLM-centric autonomous agents suffer from three intractable failure modes: non-determinism, operational hallucination, and untraceable execution. These properties disqualify LLM-centric agents from high-criticality, regulated, or auditable deployments. We present HYDRA-Ω (Hybrid Domain-partitioned Reasoning Agent), a multi-agent architecture that resolves these failures through a single design inversion: the LLM operates as a typed sensor only, parsing unstructured input into schema-validated BeliefDeltas, while a BDI-HTN cognitive engine performs all reasoning and plan selection over a statically defined, offline-verified plan library. Contributions: (C1) A six-step BDI-HTN cognitive cycle (Perceive--Believe--Desire--Plan--Execute--Learn) with precisely bounded component interfaces. (C2) A typed LLM-Sensor with a proof that structural hallucination is impossible. (C3) A personality vector P=(rho, alpha, sigma, tau) integrated into the HTN scoring function. (C4) Episodic anti-pattern memory with live TTL and confidence EMA. (C5) Safety by construction: the plan library is PDDL-verified offline. We prove four formal properties and demonstrate all guarantees empirically on a three-CVE benchmark.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo