Quickstart
Install agent-observability-sdk and point your Python agent at the dashboard. Covers the agent-transport path (observability auto-emits) and where to go for raw LiveKit workers.
Install
pip install agent-observability-sdklivekit-agents>=1.5.2,<1.6, pytest>=7.0, and httpx>=0.24 are hard deps and
installed automatically. Python ≥ 3.10.
Point it at the dashboard
Every upload path reads the same environment. The URL is required; the agent id is strongly recommended so sessions are parented to an agent row on the dashboard.
export AGENT_OBSERVABILITY_URL=https://obs.example.com
export AGENT_OBSERVABILITY_AGENT_ID=9c2f7e3d-4b8a-4d2e-9f1b-…See the full variable list in the Client SDKs overview.
agent-transport workers
If your agent runs on agent-transport's AudioStreamServer, you're done —
the transport layer emits the observability tags and runs judges through its own
EvaluationConfig. You don't call the LiveKit helpers. You may still want to
pull judges from this package into your config:
from agent_observability.livekit.judges import (
default_judges,
IntentAccuracyJudge,
rigid_response_accuracy_judge,
)
from agent_transport import AudioStreamServer
from agent_transport.evaluation import EvaluationConfig
from livekit.agents.evals import accuracy_judge
ctx.evaluation = EvaluationConfig(
judge_llm=judge_llm,
judges=[
accuracy_judge(),
IntentAccuracyJudge(expected_intent="book_flight", actual_intent=...),
rigid_response_accuracy_judge(expected_response="...", llm=judge_llm),
*default_judges(llm=judge_llm),
],
)See the full judge reference.
Next steps
- Driving LiveKit Agents directly? See Raw LiveKit.
- Running evals from a test suite? See the pytest plugin.
Client SDKs
Instrument your agent to ship session reports and eval results to the observability server. Covers the two integration modes — agent-transport (auto-emits) and raw LiveKit (you wire it) — and the env-var contract both share.
Raw LiveKit
Instrument a worker that drives LiveKit Agents directly. init_observability emits the tag bundle the server expects; run_judges_on_report runs judges at end-of-call.