ANKR Swarm: Persona-Specialized Multi-Agent Orchestration with Dynamic Spawning and Universal Synthetic Data Generation

Read full text Summarise with VIRAT
Authors: Anil Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We describe two tightly coupled innovations developed in the course of building the ANKR platform — a production AI system spanning 192 registered services across logistics, maritime, compliance, and enterprise domains. The first contribution is ANKR Swarm, a persona-specialized multi-agent orchestration framework in which each agent carries a domain-specific identity (persona), domain-tuned system prompt, and curated tool access set — in contrast to systems that assign agents generic roles. ANKR Swarm implements three distinct inter-agent communication modalities — shared-context memory, direct message-passing, and blackboard coordination — and supports dynamic spawning of sub-agents at runtime based on task decomposition signals. Consensus across agent outputs is built through six configurable strategies: majority voting, best-of-N judging, weighted merge, chain refinement, debate rounds, and mixture-of-experts routing. The second contribution is Universal Synthetic Data Generation via the abstract `SDialog` interface, through which domain packages declare their persona taxonomy, scenario definitions, dialog templates, and universal tag registrations. The generation pipeline proceeds through five stages: persona instantiation, scenario selection, dialog generation, augmentation, and quality validation. Cross-domain pattern inheritance allows a new domain package to acquire approximately 70% of its dialog patterns from a shared universal library, reducing per-domain generation cost dramatically. Together, these two systems address a fundamental problem in domain-specific AI: that generic agents hallucinate in specialist domains, and that training data for specialist agents is expensive to produce from scratch for each domain.

Fields of Study & Subjects

Multi-Agent Systems and Negotiation Artificial Intelligence Applications Speech and dialogue systems

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint