Chat span emitted even when MAF instrumentation is disabled

[singankit]

Summary

When using microsoft-opentelemetry distro with MAF (Microsoft Agent Framework) instrumentation disabled, a chat gpt-4.1 span is still being emitted. This span should not appear when MAF instrumentation is turned off.

Repro Steps

1. Use microsoft-opentelemetry distro
2. Disable MAF instrumentation
3. Invoke an agent
4. Observe that a chat span still shows up under the invoke_agent trace

Expected Behavior

No chat span should appear when MAF instrumentation is disabled.

Actual Behavior

A chat gpt-4.1 span (7.2s, 69 tokens) is emitted with full gen_ai attributes (input/output messages, token usage, cost, agent name/id). The span includes attributes like:

• gen_ai.agent.name
• gen_ai.agent.id
• microsoft.gen_ai.cost.input
• microsoft.gen_ai.cost.output
• gen_ai.usage.input_tokens
• gen_ai.usage.output_tokens
• gen_ai.response.model: gpt-4.1

Screenshot

(See attached image showing the chat span with metadata appearing under invoke_agent trace even with MAF instrumentation disabled)

Environment

• Distro: microsoft-opentelemetry (Python)
• Agent Framework: MAF (Microsoft Agent Framework)
• Model: gpt-4.1

Related

[rads-1996]

@singankit Can you share your application code? Can you also share your distro setup?

[rads-1996]

@singankit Reminder: Can you share your application code? Can you also share your distro setup?