Artifacts are not processed by the subagent through prompt instructions - Guidance required

[subbu4github]

I am having a coordinator agent, which will ask series of question and ask the user to upload files based on the user response. I am using save artifacts as plugin, so as and when the user uploads the file its getting created as artifacts. I am storing the name of the artifacts in the state variable. Based on the input I am transferring the control the subagent who can answer the user query. I have given the load_artifacts tool to the sub agent. In the prompt instruction, I have mentioned that """read the artifact {state variable name}. analyse the content and address the user queries....""". But the subagent is not reading the artifact and not producing any output. I tried switch it from subagent to AgentTool. Even in that case the result is same. How to address this issue. Any guidance in this regard is much appreciated. Thanks.

[agrovr]

LoadArtifactsTool does not load a file merely because its name appears in the instruction. The model must first emit a load_artifacts tool call with an artifact_names array; ADK then appends those contents to the following model request. The content is temporary and is not written back into session history.

A minimal sub-agent should look like this:

from google.adk.agents import LlmAgent
from google.adk.tools.load_artifacts_tool import LoadArtifactsTool

reader = LlmAgent(
    name="artifact_reader",
    model="gemini-2.5-flash",
    instruction=(
        "Before answering any question about an uploaded file, call "
        "load_artifacts with the exact available filename. Only answer "
        "after the artifact content has been loaded."
    ),
    tools=[LoadArtifactsTool()],
)

The Runner executing the coordinator/sub-agent must also be configured with the same artifact_service. Session-scoped artifacts are keyed to the same app/user/session context. A normal transfer within that runner preserves the context; running the sub-agent under another runner/session will not find the same upload.

You do not need to maintain a separate filename list in state: LoadArtifactsTool lists available artifacts in the model instructions automatically. If you do inject a state value, use the exact saved filename and ADK's state-template syntax, for example {uploaded_artifact_name}. A prose placeholder such as {state variable name} only works if that is a real valid state key.

The quickest checks are:

• confirm the runner has an artifact_service;
• confirm list_artifacts() contains the exact filename in the sub-agent invocation;
• confirm load_artifact(filename) returns a Part rather than None;
• confirm the artifact has the correct MIME type;
• inspect the trace and verify the model emitted load_artifacts({"artifact_names": [...]}).

If there is no tool call, strengthen the instruction as above. If there is a tool call but no content, it is an artifact scope/name/service issue rather than a prompting issue.

Sources:

• ADK artifact guide: https://adk.dev/artifacts/#using-loadartifactstool
• Current implementation: https://github.com/google/adk-python/blob/main/src/google/adk/tools/load_artifacts_tool.py