Async agent

Use this for the standard case: an async agent function that awaits an LLM call and returns a single response.

TypeScript
Python

import { registerOTel, wrapAgent } from "@uselemma/tracing";

registerOTel();

const wrapped = wrapAgent("my-agent", async ({ onComplete }, input: { userMessage: string }) => {
  const response = await callLLM(input.userMessage);
  onComplete(response);
  return response;
});

const { result, runId } = await wrapped({ userMessage });

onComplete(response) records the output explicitly. You can also omit it and rely on the return value being captured automatically:

const wrapped = wrapAgent("my-agent", async (_ctx, input: { userMessage: string }) => {
  return await callLLM(input.userMessage);
});

from uselemma_tracing import register_otel, TraceContext, wrap_agent

register_otel()

async def run_agent(ctx: TraceContext, user_message: str) -> str:
    response = await call_llm(user_message)
    ctx.on_complete(response)
    return response

wrapped = wrap_agent("my-agent", run_agent)

result, run_id, _ = await wrapped(user_message)

ctx.on_complete(response) records the output explicitly. You can also omit it and rely on the return value being captured automatically:

async def run_agent(ctx: TraceContext, user_message: str) -> str:
    return await call_llm(user_message)  # return value captured automatically as output

To add per-call LLM visibility (prompt, response, tokens), register the matching OpenInference instrumentor for your provider at startup. Details: Provider instrumentation.

Metrics

Tracing