Wicked AI Observatory (WAO): A Multi-Agent System for Dynamic Mapping of Wicked Problems in Artificial Intelligence — Architecture and Setup

Wicked AI Observatory (WAO) is a multi-agent system that continuously monitors, classifies, and maps AI governance challenges as a dynamic knowledge graph. AI governance challenges are paradigmatic wicked problems in the sense of Rittel and Webber (1973): they resist definitive formulation, involve stakeholders with conflicting goals, evolve over time, and have no solutions that are unambiguously correct. WAO operationalises this insight as a working monitoring infrastructure. This Technical Report (Version 0.7) documents the architecture, agent design, installation procedure, and current empirical state of the system as of April 2026. It is intended as a companion to the source code and as a citable reference for users and contributors. System overview. Seven specialised agents (Monitor, Classify, Wickedness Analysis, Evaluation, Proposal, Human Analyst, Explanation) coordinate through a shared Neo4j knowledge graph. The Monitor Agent ingests events from 13 RSS feeds and a Zotero group library; the Classify Agent applies LLM-based relevance filtering across a registry of 29 factors; the Wickedness Agent evaluates each factor against a 14-parameter wickedness vector derived from Rittel & Webber (1973). Human-in-the-loop calibration ensures normative weighting decisions remain transparent and auditable. Current state (April 2026). The graph contains 29 factors across 13 dimensions (grouped into Institutional, Societal, and Planetary clusters), 11 stakeholders, and 172 events ingested across five monitoring cycles. The five highest wickedness scores are observed for Compressed Kill Chains (96), Regulatory Capture (91), Accountability Gaps (90), Autonomous Weapon Systems (90), and Alignment Failures (88). Reproducibility. The full source code, the factor registry, and a Cypher export of the initial knowledge graph are openly available in the accompanying repository at github.com/pholubar/wicked-ai-observatory. The project website is at wicked-ai-observatory.org. Licensing. Code is released under the MIT License; documentation, factor registry, and graph data are released under CC BY 4.0. Status. This is a working technical document for an actively developed research system. A more comprehensive treatment of the theoretical foundations, the four-level validation framework, and the empirical findings from the first monitoring cycles is in preparation as a separate publication.