Describe the bug
AgnoInstrumentor currently produces spans containing extremely verbose metadata - such as nested objects, full config dumps, which makes the trace data practically unusable for debugging or observability. The noise overwhelms key insights, making the tracing ineffective and resulting in incorrect tracing costs.

Some numbers to visualise the problem:

• over 20 000 characters of metadata in user input
• 212 total json lines in the output - 209 are metadata. The real model response was only 3 sentences.

To Reproduce
Use AgnoInstrumentor().instrument()

Expected Behaviour
Spans should show only essential parts like user message, output text. Metadata should be stored in metadata.

Actual Behaviour
Tracing UIs display spans exceeding 200 lines or more, filled with nested agent configs, state, and irrelevant fields.
Searching for the actual user message becomes extremely difficult, practically impossible.

Screenshots
Input message:

Output message:

Additional context
We are using Langfuse for tracing.