Raw LiveKit

agent-transport's AudioStreamServer wires observability internally. When you drive LiveKit Agents directly — your own AgentServer with @server.rtc_session(...) — two helpers replace the hand-rolled plumbing:

• init_observability(ctx.tagger, …) — validates the upload URL (raises if unset) and emits the full tag bundle (agent_id, account_id, agent.name, transport) the server's ingest path expects. One call instead of ~20 lines of tagger.add(...).
• run_judges_on_report(report, judges=…) — wraps the JudgeGroup construction, exception handling, structured logging, and llm.aclose() cleanup. Mix LiveKit built-ins with SDK judges freely.

Minimal worker

from agent_observability.livekit import Goal, init_observability, run_judges_on_report
from agent_observability.livekit.judges import default_judges
from livekit.agents import AgentServer, JobContext
from livekit.agents.evals import accuracy_judge

server = AgentServer()


async def on_session_end(ctx: JobContext) -> None:
    report = ctx.make_session_report()
    await run_judges_on_report(
        report,
        judges=[
            accuracy_judge(),     # LiveKit built-in
            *default_judges(),    # 4 ground-truth-free SDK judges
        ],
    )


@server.rtc_session(agent_name="support-bot", on_session_end=on_session_end)
async def entrypoint(ctx: JobContext) -> None:
    init_observability(
        ctx.tagger,
        agent_id="9c2f7e3d-…",   # stable opaque UUID
        agent_name="support-bot",
        account_id="acct-7",
        transport="text",
        goals=[
            Goal("identity-check", "Confirm the caller's identity before account changes"),
            Goal("order-resolution", "Resolve the order issue or open a support ticket"),
        ],
    )
    # …your usual AgentSession.start(...) setup

That's the whole observability surface for a raw-LiveKit worker — no hand-rolled tagger.add(...), no JudgeGroup boilerplate, no llm.aclose() cleanup.

Conversation goals

The goals parameter declares task outcomes the server's goal analyzer judges after each session — no judging code in the agent, just data. Each entry is a Goal(name, description) — both fields required:

from agent_observability.livekit import Goal

init_observability(
    ctx.tagger,
    agent_id="9c2f7e3d-…",
    goals=[
        # name + description — the description is what the LLM judge evaluates
        Goal("identity-check", "Confirm the caller's identity before account changes"),
        # descriptions may contain colons; names may not
        Goal("escalation-policy", "Escalate only after: two failed resolution attempts"),
    ],
)

A Goal validates itself at construction (and duplicate names raise before any tag is emitted):

• names are the goal's stable identity — non-empty, unique within the call, and colon-free (the wire format goal:<name>:<description> splits at the first colon).
• descriptions are required free text. Phrase them as observable conversation behavior — the judge only sees the transcript (guidance).

Verdicts (met/unmet, reasoning, what went wrong) land on the agent's Conversation Goals tab once the server has OPENAI_API_KEY set.

Running on agent-transport's AudioStreamServer (which wires the observability bootstrap internally)? Call add_goal_tags(ctx.tagger, goals) instead — same goal validation, no second bootstrap, no upload-URL check.

URL resolution

init_observability raises if neither LIVEKIT_OBSERVABILITY_URL nor the AGENT_OBSERVABILITY_URL fallback is set — there's no point continuing if the report has nowhere to go. For a non-fatal, warn-only contract, call ensure_observability_url() yourself instead.

Runnable example

A complete text-only worker (tool call, recording options, judges at end-of-call) lives in the repo at plugins/examples/python/text_only_livekit_worker.py:

export LIVEKIT_OBSERVABILITY_URL=https://obs.example.com
export AGENT_OBSERVABILITY_AGENT_ID=9c2f7e3d-4b8a-4d2e-9f1b-textonly
export OPENAI_API_KEY=sk-...
uv run plugins/examples/python/text_only_livekit_worker.py console --text --record

See the judge reference for the full catalogue.