Problem Statement
I maintain a Python LangChain/LangGraph codebase, and after working on the equivalent JavaScript integration in PostHog/posthog-js#4556, I checked whether the Python SDK supported the same LangChain v1 middleware flow.
The current CallbackHandler works fine with create_agent, so tooday, the only way to add PostHog is when the agent runs:
agent.invoke(
{"messages": [...]},
config={"callbacks": [CallbackHandler(...)]},
)The things is that LangChain v1 also supports middleware as the main way to package integrations with an agent. PostHog does not currently support that option:
agent = create_agent(
model=model,
tools=tools,
middleware=[PostHogMiddleware(...)],
)The JavaScript SDK had the same gap, tracked in PostHog/posthog-js#2907 and implemented in PostHog/posthog-js#4556.
Solution Brainstorm
Add an optional LangChain v1 middleware alongside the existing callback integration. Both should produce the same PostHog traces, generations, and tool spans, while letting users choose the integration style that fits their agent.
Would this be useful on the Python side, or are callbacks and OpenTelemetry intentionally the preferred direction here? If middleware makes sense, I’m happy to work on it.
Problem Statement
I maintain a Python LangChain/LangGraph codebase, and after working on the equivalent JavaScript integration in PostHog/posthog-js#4556, I checked whether the Python SDK supported the same LangChain v1 middleware flow.
The current
CallbackHandlerworks fine withcreate_agent, so tooday, the only way to add PostHog is when the agent runs:The things is that LangChain v1 also supports middleware as the main way to package integrations with an agent. PostHog does not currently support that option:
The JavaScript SDK had the same gap, tracked in PostHog/posthog-js#2907 and implemented in PostHog/posthog-js#4556.
Solution Brainstorm
Add an optional LangChain v1 middleware alongside the existing callback integration. Both should produce the same PostHog traces, generations, and tool spans, while letting users choose the integration style that fits their agent.
Would this be useful on the Python side, or are callbacks and OpenTelemetry intentionally the preferred direction here? If middleware makes sense, I’m happy to work on it.