Paper Details Low-Level Architecture For Spatial Agentic AI

Geospatial agentic AI, the practice of letting autonomous AI agents plan and act on mapping, sensing and location data, has mostly existed as a set of high-level diagrams describing what the agents should do, not how to build them. The paper filed on arXiv on October 9, 2026, by Constantinos Papantoniou and Brian Hilton tries to close that gap for one specific framework, called ANASSA, which an earlier paper had set out as a high-level orchestration model for governed spatial agent execution.

The new work applies a method called Design Science Research to turn that high-level model into something closer to a build spec. Instead of naming only broad components such as “orchestration” or “governance”, the authors decompose those responsibilities into concrete architectural pieces: 47 logical roles an agent system can fill, versioned task graphs that record how a job moved through the system, capability-based selection for choosing which agent handles a step, typed contracts between components, explicit runtime states, spatial validation checks, governance controls, failure and recovery paths, and provenance records that trace where a result came from. The authors demonstrate the design with an end-to-end execution trace and evaluate it against four criteria: requirements traceability, internal consistency, responsibility preservation and whether the design stays independent of any particular AI model, mapping platform or deployment setup.

The stated goal is a bridge between the original ANASSA framework and whoever eventually implements it, so that a governed spatial agent workflow can be inspected, rebuilt and tested without committing to one vendor’s model, protocol or cloud environment ahead of time.

The numbers

• The architecture specifies 47 logical roles an agent in the system can occupy, covering orchestration, validation, governance and recovery duties separately rather than as one catch-all agent.
• The paper runs 14 pages, with 13 tables carrying the decomposition detail and 2 figures illustrating the architecture and its execution trace.

For anyone deciding whether to build a spatial agent pipeline on a published framework rather than a proprietary one, the practical question is usually whether the framework specifies enough to code against without reverse-engineering someone else’s assumptions. A high-level framework alone leaves that work undone; naming 47 distinct roles and typed contracts between them gives an implementer something closer to an interface list than a mission statement, which changes the build decision from “interpret this diagram” to “implement this contract.” Whether other teams adopt ANASSA specifically, the paper is a data point in how academic agentic-AI work is moving from architecture diagrams toward specifications detailed enough to hand to an engineering team.