Decorate a function with @multi_asset(outs={...}) where each key in outs is an AssetOut defining a separate asset with its own key, description, and optional io_manager_key
Inside the function body, yield Output(value, output_name=key) for each asset output, allowing Dagster to track each output as an independent asset with its own materialization record
Define internal_asset_deps={output_key: {AssetKey('other_output_key')}} in @multi_asset to declare which outputs depend on which other outputs within the same function, enabling Dagster to build the correct asset graph
Use can_subset=True if the multi-asset can be partially executed; implement the context.selected_output_names check to skip computation for non-selected outputs
Assign different io_manager_key values to different outputs if each asset should be persisted to a different storage backend
Known gotchas
If internal_asset_deps is not specified, Dagster assumes all outputs depend on all inputs, which can cause unnecessary re-materialization of unrelated assets when only one upstream changes
Yielding outputs in a different order than declared in outs does not cause an error but can confuse lineage tracking; always yield outputs by their declared name
can_subset=True requires explicit logic to skip unneeded computation; if the function always runs all computations regardless of selected_output_names, partial execution wastes resources without surfacing an error
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp