rfc(decision): Prevent duplicate AI Client Spans by alexander-alderman-webb · Pull Request #155 · getsentry/rfcs

alexander-alderman-webb · 2026-03-09T15:15:40Z

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

alexander-alderman-webb · 2026-03-09T15:15:40Z

Rendered RFC

nicohrubec · 2026-03-10T09:42:11Z

+
+Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.
+
+Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


Might be helpful here to give an example of what you mean by biased statistics (e.g. double counting of tokens)

nicohrubec · 2026-03-10T09:52:10Z

+
+#### Pros
+
+More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.


Suggested change
More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
nit but not sure if this is universally true so I'd phrase this less definitive

nicohrubec · 2026-03-10T09:54:31Z

+ └── API call
+```
+
+Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.


Suggested change
Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

nicohrubec · 2026-03-10T09:57:56Z

+Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.
+
+Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.
+


From our (js) side the most important drawback is that this solution doesn't work for our langchain use case

nicohrubec · 2026-03-10T10:00:45Z

+
+#### Cons
+
+The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.


This is only true for the "AI Client Scope" case right? For the boolean case I think this is only marginally more complex than the parent solution

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes

rfc(decision): Prevent duplicate AI Client Spans
e2d49ed

alexander-alderman-webb force-pushed the rfc/prevent-duplicate-ai-client-spans branch from 6f90149 to e2d49edCompare March 9, 2026 15:15

alexander-alderman-webb added 2 commits March 9, 2026 16:27

.
81442ad

.
e7c7b03

alexander-alderman-webb requested review from RulaKhaled, andreiborza, ericapisani and nicohrubec March 10, 2026 08:06

alexander-alderman-webb marked this pull request as ready for review March 10, 2026 08:06

nicohrubec reviewed Mar 10, 2026
View reviewed changes


		Many statistics displayed in the AI Agents Insights module are based on AI Client Spans. Statistics, including the total incurred cost, require all model calls to be instrumented with AI Client Spans exactly once. Model calls initiated by agents are always instrumented in agent libraries to support users of an agent library with a model provider not yet supported by the corresponding SDK. Conversely, "low-level" libraries of model providers are also instrumented with AI Client Spans to support users who call models outside of a supported agent framework.

		Users of a supported agent framework with a supported model provider therefore generate two AI Client Spans, resulting in biased statistics in the AI Agents Insights module.


		#### Pros

		More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.

	More information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.
	In some cases more information used to populate the attributes of AI Client Spans is available in the client library than in the agent library abstraction.

	Some data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.
	In some cases data exposed by the client library is not available at the agent's model abstraction level. For example, the concrete response model used to populate the `gen_ai.response.model` attribute is always available in the `openai` package, but never returned by the `Model` abstraction used by agents in the `openai-agents` package.

		Disabling spans at the client library level results in a loss of information when the agent library's abstractions expose less information about model calls, and there is no mechanism to capture information from the client library.

		Additionally, since the model abstraction used by an agent may contain user-defined code, there can be spans in between the AI Client Spans for the agent and client library. The parent-aware disabling only applies if there are no spans in between AI Client Spans in the trace hierarchy.


		#### Cons

		The solution introduces complexity to the SDK. An integration would mutate the attributes set on a span created by another integration when the agent and client libraries are both instrumented.

Uh oh!

Uh oh!

Conversation

alexander-alderman-webb commented Mar 9, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nicohrubecMar 10, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

alexander-alderman-webb commented Mar 9, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nicohrubecMar 10, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

alexander-alderman-webb commented Mar 9, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nicohrubecMar 10, 2026• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

alexander-alderman-webb commented Mar 9, 2026•
edited
Loading

nicohrubecMar 10, 2026•
edited
Loading

alexander-alderman-webb commented Mar 9, 2026•
edited
Loading

nicohrubecMar 10, 2026•
edited
Loading

alexander-alderman-webb commented Mar 9, 2026•
edited
Loading

nicohrubecMar 10, 2026•
edited
Loading

alexander-alderman-webb commented Mar 9, 2026•
edited
Loading

nicohrubecMar 10, 2026•
edited
Loading

alexander-alderman-webb commented Mar 9, 2026•
edited
Loading