[BUG] OTEL span reports accumulated_usage instead of per-invocation usage, causing inflated token metrics in Langfuse

[afarntrog]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

v1.33

Python Version

3.12

Operating System

macox

Installation Method

pip

Steps to Reproduce

• Install Strands Python SDK via pip

• Configure OTEL env vars to export traces to Langfuse:

  • OTEL_EXPORTER_OTLP_ENDPOINT
  • OTEL_EXPORTER_OTLP_HEADERS
  • OTEL_EXPORTER_OTLP_PROTOCOL
  • DISABLE_ADOT_OBSERVABILITY=true
  • OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_tool_definitions

• Run an agent across multiple invocations in a multi-chat session

• Observe token usage reported in Langfuse per span

Expected Behavior

Each OTEL span should report token usage for that specific invocation only. e.g., in a session with 10 invocations of ~100k tokens each:

• Request 1 → 100k tokens
• Request 2 → 100k tokens
• Request 3 → 100k tokens

Total: ~1M tokens

Actual Behavior

Each OTEL span reports the session-lifetime accumulated_usage instead of per-invocation usage. e.g.:

• Request 1 → 100k tokens ✅
• Request 2 → 200k tokens ❌ (should be 100k)
• Request 3 → 300k tokens ❌ (should be 100k)

Langfuse then sums these, resulting in wildly inflated token counts and cost estimates. Additionally, reset_usage_metrics() does NOT reset accumulated_usage — this appears to be intentional per the test suite, but means there is no workaround.

Additional Context

Root cause identified in src/strands/telemetry/tracer.py: The OTEL span reporter explicitly uses accumulated_usage when setting span attributes:

accumulated_usage = response.metrics.accumulated_usage
attributes.update({
    "gen_ai.usage.prompt_tokens": accumulated_usage["inputTokens"],
    "gen_ai.usage.input_tokens":  accumulated_usage["inputTokens"],
    "gen_ai.usage.output_tokens": accumulated_usage["outputTokens"],
    ...
})

In metrics.py, reset_usage_metrics() is intentionally designed to NOT clear accumulated_usage:

# Verify accumulated_usage is NOT cleared
assert event_loop_metrics.accumulated_usage["inputTokens"] == 11

Possible Solution

In tracer.py, replace accumulated_usage with the latest invocation's usage:

agent_invocations[-1].usage

instead of:

response.metrics.accumulated_usage

This would report only the current invocation's token usage on each span.

Related Issues

No response

[en-yao]

Hi, I'd like to work on this.

I've confirmed the root cause — end_agent_span in tracer.py reads response.metrics.accumulated_usage (session-lifetime cumulative) instead of the current invocation's usage. The per-invocation usage already exists at response.metrics.latest_agent_invocation.usage, populated by update_usage() in metrics.py.

My proposed fix is a one-liner replacing accumulated_usage with latest_agent_invocation.usage. I have it working locally with tests passing.

[afarntrog]

Hi @en-yao, that would be great! Thanks!

[JackYPCOnline]

@poshinchen What is your thought?

[Zelys-DFKH]

Confirmed the root cause. end_agent_span in tracer.py reads response.metrics.accumulated_usage, which grows with every request in a session. The fix is a one-liner: use response.metrics.latest_agent_invocation.usage instead, with accumulated_usage as the fallback when no invocation is recorded. Happy to open a PR.

[poshinchen]

Why do we have two PRs lol?

[poshinchen]

@en-yao After discussing with the team, we decided to allow the user to opt in to using tokens from the current turn.

Similar to adding full tool_definitions into the span, we will use the flag gen_ai_use_latest_invocation_tokens to switch between latest_agent_invocation and accumulated_usage.

[Zelys-DFKH]

Saw #2017 was already in progress, so my comment was redundant. Good call on the flag; keeping accumulated_usage as the default means nobody's dashboards break silently.

[poshinchen]

@en-yao Left the comment in your PR, after that I'll be fine to approve