Finding Slow Tool Calls in Vercel AI SDK v7 Telemetry

A lot of AI demos look incredibly fast until you run them in production. My weather-checking agent worked perfectly locally, but users complained it was freezing for ten seconds before typing a response.

Before Vercel AI SDK v7, debugging latency meant littering my code with console.time(). The new @ai-sdk/otel package fixes this by hooking directly into Node.js tracing channels.

Wiring up OpenTelemetry

You have to configure an OpenTelemetry provider first. I exported my traces to a local Jaeger instance to visualize them. The SDK handles the rest automatically once you import the instrumentation module.

// instrumentation.ts
import { registerOTel } from '@vercel/otel';
import { AISDKExporter } from '@ai-sdk/otel';

export function register() {
  registerOTel({
    serviceName: 'weather-agent',
    traceExporter: new AISDKExporter(),
  });
}

// agent.ts
import { streamText } from 'ai';
import { openai } from '@ai-sdk/openai';
import { z } from 'zod';

export async function POST(req: Request) {
  return streamText({
    model: openai('gpt-4o'),
    prompt: 'What is the weather in London?',
    tools: {
      getWeather: {
        description: 'Get current weather',
        parameters: z.object({ city: z.string() }),
        execute: async ({ city }) => {
          // This API call was the culprit
          const res = await fetch(`https://api.slow-weather.com/v1/${city}`);
          return res.json();
        },
      },
    },
  });
}

Reading the Traces

When I checked the Jaeger dashboard, the problem was obvious. The trace showed the initial LLM call took 800 milliseconds. Then there was a massive eight-second block labeled tool-execution: getWeather. The final LLM call to format the response took another 600 milliseconds.

The model was not slow. The external weather API I was using had terrible latency.

The v7 observability features give you a granular breakdown of every step: how long the model thought, how long the tool ran, and how long the stream took to close. If you are building agents that rely on third-party APIs, this level of tracing is not optional. You have to know what is actually slowing you down.