BeeAI tracing with Future AGI

Send BeeAI traces to an Observe project using BeeAI’s recommended OpenInference instrumentor and the Future AGI tracer provider. You do not need traceai-beeai.

Install

Use Python 3.11–3.13. This example uses these package versions:

pip install "beeai-framework==0.1.85" "openinference-instrumentation-beeai==0.1.31" "fi-instrumentation-otel==1.2.2"

Set your credentials

Create an API key and secret key in Future AGI, then set them in your application’s environment:

export FI_API_KEY="your-futureagi-api-key"
export FI_SECRET_KEY="your-futureagi-secret-key"
# For a self-hosted deployment, set the origin, without /tracer/v1/traces:
# export FI_BASE_URL="https://your-futureagi-host"

The default destination is Future AGI Cloud. register() configures OTLP/HTTP trace export and authentication; no separate exporter or global provider is needed.

Trace a tool

Save this as beeai_example.py. The example runs a local tool without calling a model provider. Configure instrumentation once, before running BeeAI operations.

import asyncio

from fi_instrumentation import register
from fi_instrumentation.fi_types import ProjectType
from fi_instrumentation.otel import SemanticConvention
from openinference.instrumentation import TraceConfig, using_attributes
from openinference.instrumentation.beeai import BeeAIInstrumentor

provider = register(
    project_name="beeai-agent",
    project_type=ProjectType.OBSERVE,
    semantic_convention=SemanticConvention.OPENINFERENCE,
)
instrumentor = BeeAIInstrumentor()
instrumentor.instrument(
    tracer_provider=provider,
    config=TraceConfig(hide_inputs=True, hide_outputs=True),
)

from beeai_framework.tools import tool


@tool
def shipping_days(region: str) -> str:
    """Estimate delivery time for a shipping region."""
    return "3 business days" if region == "domestic" else "7 business days"


async def main():
    with using_attributes(session_id="shipping-session", user_id="example-user"):
        result = await shipping_days.run({"region": "domestic"})
        print(result.get_text_content())


try:
    asyncio.run(main())
finally:
    provider.force_flush()
    instrumentor.uninstrument()
    provider.shutdown()

python beeai_example.py

In Future AGI, open Observe, select beeai-agent, and open the shipping_days tool trace. Model calls made through BeeAI’s ChatModel are also instrumented; configure your model provider separately when you add one.

SemanticConvention.OPENINFERENCE tells Future AGI how to interpret these spans. Keep it explicit rather than using the default Future AGI attribute convention.

The example hides input and output content. To capture it, change both TraceConfig options to False after reviewing your data policy. Hiding content does not sanitize every metadata field or exception message; avoid putting secrets in session IDs, user IDs, metadata, or errors.

Migrate from traceai-beeai

Replace the wrapper with the setup above, then remove traceai-beeai from your dependencies. Do not enable both instrumentors for the same BeeAI operations.

Previous API	Replacement
configure_beeai_tracing(tracer_provider=provider)	BeeAIInstrumentor().instrument(tracer_provider=provider)
BeeAIInstrumentorWrapper	openinference.instrumentation.beeai.BeeAIInstrumentor
Endpoint-only configure_beeai_tracing(...)	Create a provider with register() first; set FI_BASE_URL for a self-hosted Future AGI origin.
setup_traceai_exporter(...) from traceai_beeai._instrumentor	Use register() for Future AGI export. For a custom OTLP destination, pass a standard OTLPSpanExporter as register(span_exporter=...), with your endpoint and headers.
Middleware session_id / user_id	Wrap the operation in openinference.instrumentation.using_attributes(...), as above.

Existing traceai-beeai==0.1.0 installations remain available on PyPI. If your application calls TraceAIMiddleware.on_*, create_tracing_middleware(), or attribute helpers directly, keep that pin until you replace those calls with standard OpenTelemetry spans or BeeAI’s upstream instrumentation. These custom callbacks are not a drop-in match for OpenInference events. Preserve any custom attributes, exporter settings, and dashboards you depend on before uninstalling the package.

Troubleshooting

• No spans: use the package versions above, call .instrument() before executing tools or model calls, and pass the provider returned by register(). Setting an OTLP endpoint environment variable alone does not create an exporter.
• Unexpected span fields: confirm the provider uses SemanticConvention.OPENINFERENCE.
• Duplicate spans: remove the old wrapper and custom tracing middleware for operations now handled upstream. Do not add another model instrumentor for the same BeeAI model calls.
• Missing final traces: flush and shut down the provider when a short-lived process exits, as in the example.