feat(google-adk): add ADK instrumentation and e2e coverage

[Abhijeet Prasad (AbhiPrasad)]

resolves #1477

Add Braintrust instrumentation for @google/adk with auto-instrumentation support by default when users run with Braintrust's loader hook.

This adds tracing for:

• Runner.runAsync
• BaseAgent.runAsync and subclasses
• FunctionTool.runAsync

It also adds Google ADK e2e coverage for both hook-driven and wrapped execution.

Because ADK uses @google/genai under the hood, underlying LLM spans continue to come from the existing Google GenAI instrumentation.

Recommended usage: auto-instrumentation

The primary way users should adopt this is via Braintrust auto-instrumentation:

node --import braintrust/hook.mjs app.mjs

Then they can use @google/adk normally and Braintrust will trace supported ADK operations automatically.

Example:

import * as adk from "@google/adk";
import { initLogger } from "braintrust";

initLogger({ projectName: "my-google-adk-app" });

const agent = new adk.LlmAgent({
  name: "weather_agent",
  model: "gemini-2.5-flash-lite",
  instruction: "Answer weather questions",
});

const runner = new adk.InMemoryRunner({ agent, appName: "weather-app" });

await runner.sessionService.createSession({
  appName: runner.appName,
  userId: "user-1",
  sessionId: "session-1",
});

for await (const event of runner.runAsync({
  userId: "user-1",
  sessionId: "session-1",
  newMessage: {
    role: "user",
    parts: [{ text: "What's the weather in Paris?" }],
  },
})) {
  console.log(event);
}

If needed, it can be disabled with:

BRAINTRUST_DISABLE_INSTRUMENTATION=google-adk node --import braintrust/hook.mjs app.mjs

Manual usage

Manual wrapping is also available for cases where users want explicit control:

import * as adk from "@google/adk";
import { wrapGoogleADK } from "braintrust";

const wrapped = wrapGoogleADK(adk);

Screenshot

[Abhijeet Prasad (AbhiPrasad)]

wait this need some more work. I realized the parent child relationships are wrong 😢

[Luca Forstner (lforst)]

wait this need some more work. I realized the parent child relationships are wrong 😢

I honestly don't know why this is so painful in this SDK. Tracing channels make it harder because they are event based. It should honestly just be a simple wrapping call imo.

[Abhijeet Prasad (AbhiPrasad)]

Okay fixed the parent-child relationships! This is good to review now.

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	next@​15.5.15 ⏵ 14.2.34	+20	-30	+4	-46
	@​ai-sdk/​openai@​3.0.52 ⏵ 1.3.24	-1		-3
	@​ai-sdk/​openai@​3.0.52 ⏵ 2.0.53	-1		-3
	@​opentelemetry/​exporter-trace-otlp-http@​0.53.0 ⏵ 0.200.0
	@​opentelemetry/​exporter-trace-otlp-http@​0.53.0 ⏵ 0.205.0
	@​opentelemetry/​context-async-hooks@​1.26.0 ⏵ 2.6.1			-4
	openai@​5.23.2
	@​vercel/​otel@​0.1.0
	cohere-ai@​7.21.0
	cohere-ai@​8.0.0
	@​types/​react-dom@​18.3.7
	@​braintrust/​otel@​0.1.2
	@​opentelemetry/​sdk-trace-base@​1.26.0 ⏵ 2.0.0	+1
	cohere-ai@​7.20.0
	@​ai-sdk/​openai@​3.0.52 ⏵ 0.0.62	+28		-5	-1
	@​opentelemetry/​otlp-exporter-base@​0.53.0 ⏵ 0.205.0	+1		-1
	react@​19.2.5 ⏵ 18.3.1			+1
	@​opentelemetry/​otlp-exporter-base@​0.53.0 ⏵ 0.35.0	+1			-3
	@​ai-sdk/​openai@​3.0.52 ⏵ 2.0.57	+28		-3
	@​ai-sdk/​openai@​3.0.52 ⏵ 3.0.7	+27		-2
	@​opentelemetry/​otlp-transformer@​0.53.0 ⏵ 0.35.0	+1		-3	-3
	@​vercel/​otel@​2.1.0
	@​opentelemetry/​context-async-hooks@​1.26.0 ⏵ 1.9.0	+27		+1	-5
	@​opentelemetry/​sdk-trace-base@​1.26.0 ⏵ 1.9.0	+22			-5
	zod@​4.1.11
	ai@​6.0.159 ⏵ 5.0.76	-10
	typescript@​5.4.4
	braintrust@​2.2.2 ⏵ 2.2.1-rc.0	+1
	@​opentelemetry/​exporter-trace-otlp-http@​0.53.0 ⏵ 0.35.0	+26			-3
	react-dom@​19.2.5 ⏵ 18.3.1	-7
	ai@​6.0.159 ⏵ 6.0.1	-8
	cohere-ai@​7.14.0
See 15 more rows in the dashboard

View full report