GIS teams are being asked to turn more open data into better decisions, faster. At the same time, many AI demonstrations for geospatial work still look like generic chatbots: impressive in a demo, but weak on provenance, reproducibility, map-based evidence, and human review. This talk presents a practical alternative: an open GIS agent workflow built around Hermes Agent and free/open geospatial tools.

Hermes is an open-source agent framework that supports tool use, persistent skills, profile separation, scheduled tasks, and messaging-based human supervision. In this workflow, Hermes acts as an orchestrator while a specialist GIS profile handles spatial work. The orchestrator scopes the request, keeps approval boundaries clear, and synthesizes the final brief. The GIS profile reads project context, fetches or inspects open spatial data, runs repeatable checks, prepares map-ready outputs, and returns source-backed findings for human review.

The talk will use a Sacramento-inspired public-data scenario aligned with FOSS4G NA 2026 themes of resilience and innovation. The example workflow turns open civic or environmental data into a “spatial owner brief”: a concise, map-backed summary of what changed, where issues cluster, which records look stale or incomplete, what assumptions were made, and what decisions require a human. The emphasis is not on replacing GIS professionals. It is on reducing repetitive analyst work while making AI-assisted geospatial outputs easier to inspect, reproduce, and challenge.

The proposed stack is intentionally FOSS4G-centered. The workflow can use QGIS, GDAL/OGR, Python, PostGIS or GeoPackage, GeoJSON, PMTiles, and MapLibre or OpenLayers for web visualization. Hermes provides the agent layer around those tools: profile separation, reusable GIS skills, source-aware summaries, and approval gates. The agent does not silently edit authoritative data, publish maps, or email findings. By default, it operates read-only, cites its sources, records caveats, and asks for human approval before external writes or publication.

Attendees will see how to structure a GIS agent so it remains useful and bounded: separate orchestration from spatial analysis, keep durable GIS procedures in skills instead of ad hoc prompts, require source links for claims, generate artifacts that can be opened outside the chat, and treat external datasets as untrusted until inspected. The session will also cover practical failure modes, including context pollution, role drift, stale memory, brittle demos, and overbroad permissions.

This is an application-focused talk for GIS analysts, civic technologists, open-source geospatial developers, and public-sector teams exploring GeoAI without giving up reproducibility or professional judgment. Attendees will leave with a reusable architecture, a map-brief pattern, and a checklist for adapting the approach to their own open-data portals, resilience workflows, and public decision-support projects.