Routing traces

[vitalik]

Question

What's the best way to route tracing destination for different agents?

I'm using Langfuse for tracing

initialization insturction is the following

from pydantic_ai.agent import Agent
 
Agent.instrument_all()

but my goal is that if I have two agents - I want one to trace to one server, and 2nd to other

agent1 = Agent('gpt-4o')  # This should send traces to https://langfuseserver1.com/
agent2 = Agent('somelocalmodel')  # This should send traces to https://langfuseserver2.com/

[DouweM]

@vitalik If you don't want to use the logfire SDK (which is not Pydantic Logfire-specific but really a generic OTel SDK), you can set a specific Agent's instrument property to an InstrumentationSettings object, which can be initialized with a custom tracer_provider, which dictates where the traces will actually be sent. I don't know if Langfuse's SDK lets you set up connections to different servers and get their tracer provider.

If you are OK using logfire, you can create separate Logfire instances for each server, and instrument the agents with the appropriate instance:

import logfire

logfire1 = logfire.configure(local=True, advanced=logfire.AdvancedOptions(base_url='https://langfuseserver1.com/'))
logfire2 = logfire.configure(local=True, advanced=logfire.AdvancedOptions(base_url='https://langfuseserver2.com/'))

logfire1.instrument_pydantic_ai(agent1)
logfire2.instrument_pydantic_ai(agent2)

Let me know if that doesn't work right and I'll reopen the issue!

[vitalik]

@DouweM

Could you explain what magic happens under Agent.instrument_all() or instrument=True

like why this code works out of the box ? and is there a way to add some "interceptor"

from pydantic_ai import Agent, RunContext
from langfuse import Langfuse

langfuse = Langfuse(
    secret_key="sk-lf-xxx",
    public_key="pk-lf-yyy,
    host="http://localhost:3000",
)




Agent.instrument_all()

agent = Agent(
    'openai:gpt-4o',
    deps_type=int,
    system_prompt=(
        'Use the `roulette_wheel` function to see if the ' 'customer has won based on the number they provide.'
    ),
    instrument=True,
)

how does pydantic ai knows that there is a langfuse somewhere is they are not passed to each other ?

[DouweM]

@vitalik Langfuse and Logfire (used by Pydantic AI's instrumentation features) both use the opentelemetry package, so I suppose Langfuse is registering itself as the default tracer provider, and Pydantic AI/Logfire is writing there by default.

and is there a way to add some "interceptor"

Check out span processors: https://logfire.pydantic.dev/docs/reference/api/logfire/#logfire.configure(additional_span_processors)