ArcGIS Assistant (JS 5.0 Beta) — LLM Responsibility: Orchestrator vs Custom Agent

[AduriSreenivas]

I’m evaluating arcgis-assistant with a custom arcgis-assistant-agent for a governed enterprise deployment and would like clarification on the intended LLM boundary.

From the documentation:

• The Assistant runtime includes an internal orchestrator for intent detection and agent routing.

• Custom agents define their own LLM-driven workflows inside createGraph().

In production scenarios where:

• All LLM calls must be executed server-side

• Tool planning must be auditable

• Tool execution must respect ArcGIS role-based permissions

• Destructive operations require governance controls

I would like to understand the intended separation of responsibility between:

• Assistant orchestrator (intent detection)

• Agent reasoning (planning and tool selection)

• Tool execution (FeatureLayer edits, GP services, REST operations)

Specifically:

1. Are the orchestrator LLM and the custom agent LLM independent and separately configurable?

2. Is the expected enterprise pattern to allow LLM reasoning within the Assistant runtime, or to externalize all orchestration to a backend service?

Clarification on the recommended architectural pattern for production deployments would be appreciated.

Thank you

Aduri Sreenivas