Skip to main content
Origin works with any trace you can produce. There are three on-ramps depending on what you already have.

​
1. Turnkey — live pipeline

The recommended starting point. Give Origin your pipeline and it handles tracing and scoring automatically. Your pipeline only needs @span decorators from aevyra_witness.runtime: 
from aevyra_witness.runtime import span
from aevyra_origin import diagnose_pipeline
from aevyra_origin.llm import anthropic_llm

@span("classify")
def classify(text): ...

@span("retrieve")
def retrieve(topic): ...

@span("answer", optimize=True, prompt_id="answer_v1")
def answer(q, docs): ...

def my_agent(q):
    topic = classify(q)
    return answer(q, retrieve(topic))

result = diagnose_pipeline(
    my_agent, "how do I get a refund?",
    judge=my_judge,
    rubric="Accurate and grounded in policy.",
    llm=anthropic_llm(),
)
See Quick start for the full example with a Verdict judge.

​
2. Adapter — existing framework logs

Already emitting telemetry from another framework? Parse the logs into an AgentTrace and hand it directly to Origin.

​
OpenClaw JSONL

OpenClaw streams telemetry as JSONL — one event per line. The from_openclaw_jsonl adapter handles start/end pairing, auto-parents tool calls via tool_call_id, and covers all OpenClaw event families including Task Brain (task.*, cron.*, acp.*): 
from pathlib import Path
from aevyra_witness.adapters import from_openclaw_jsonl
from aevyra_origin import Origin
from aevyra_origin.llm import anthropic_llm

lines = Path("run_2026_04_21.jsonl").read_text().splitlines()
trace = from_openclaw_jsonl(lines, ideal="expected output")

origin = Origin(llm=anthropic_llm())
result = origin.diagnose(trace=trace, score=0.4, rubric=rubric)
To mark prompts as Reflex optimization targets without annotating the event stream: 
trace = from_openclaw_jsonl(lines, optimize_prompt_ids=["planner", "responder"])

​
OpenTelemetry (LangGraph, CrewAI, AutoGen, Vercel AI SDK)

Any framework that emits OpenTelemetry spans with the GenAI semantic conventions works out of the box: 
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
from aevyra_witness.adapters import from_otel_spans
from aevyra_origin import Origin
from aevyra_origin.llm import anthropic_llm

exporter = InMemorySpanExporter()
# ... configure your OTel TracerProvider with this exporter, run your agent ...
spans = exporter.get_finished_spans()
trace = from_otel_spans(spans)

origin = Origin(llm=anthropic_llm())
result = origin.diagnose(trace=trace, score=0.4, rubric=rubric)
Plain dicts from an OTLP JSON export are also accepted by from_otel_spans.

​
Bring your own format

For structured logs from Langfuse, LangSmith, or a home-grown JSONL store, the BYO trace tutorial shows a 30-line adapter pattern that works for any source format.

​
3. Raw — AgentTrace + score

Already have an AgentTrace and a score? Use Origin.diagnose directly: 
from aevyra_origin import Origin
from aevyra_origin.llm import anthropic_llm

origin = Origin(llm=anthropic_llm())
result = origin.diagnose(
    trace=my_trace,
    score=0.4,
    rubric="Accurate, grounded in the policy docs.",
)
print(result.render())
Or via the CLI if you have the trace as a JSON file: 
aevyra-origin diagnose trace.json \
  --score 0.4 \
  --rubric rubric.txt \
  --model anthropic/claude-sonnet-4-5
AgentTrace.to_dict() / AgentTrace.to_json() serialize the trace; AgentTrace.from_dict() / AgentTrace.from_json() restore it. Non-Python producers can emit a conforming JSON object directly — see the Witness schema spec.