Observability

What	Where
Live progress	One INFO log line per step on `logging.getLogger('browser_use')`
Tokens and cost	`result.usage`, `result.cost_usd`
Every message, cell and output	`result.messages`, and the files in `agent.workspace`
Traces in your observability tool	OpenTelemetry spans from Pydantic AI

Logs

Each step logs step N: <first line of the cell>, plus the error if the cell failed, and the run ends with completed in N steps, Ns. Stream it to your UI with a logging handler: Long tasks: Live progress. BROWSER_USE_LOGGING_LEVEL=warning turns the step lines off.

Usage and cost

from browser_use import Agent

result = Agent('What is the title of https://example.com?').run_sync()
print(result.steps, result.duration_s)
print(result.usage)  # {'input': ..., 'output': ..., 'cache_read': ..., 'cache_write': ..., 'requests': ...}
print(result.cost_usd)  # model cost in USD, plus web searches; None when the price is unknown

cost_usd covers the model and web searches, not cloud browser time. See Models: Usage and cost and Output.

Transcript and workspace

result.messages is the full conversation as Pydantic AI messages: every cell the model wrote and what it returned, with secrets redacted.
agent.workspace keeps each cell’s output in .browser-use/cells/, plus screenshots and files the agent saved. It is not deleted after the run. See Files.

OpenTelemetry traces

browser-use runs on Pydantic AI, so Pydantic AI’s instrumentation traces it: one span per run, per model request (chat bu-fast, with tokens and the messages) and per tool call (execute_tool python, with the code and its output). Send them to any OpenTelemetry (OTLP) backend:

uv add opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from pydantic_ai import Agent as PydanticAgent
from pydantic_ai.models.instrumented import InstrumentationSettings

from browser_use import Agent

provider = TracerProvider()
provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))  # reads OTEL_EXPORTER_OTLP_ENDPOINT
PydanticAgent.instrument_all(InstrumentationSettings(tracer_provider=provider))

print(Agent('What is the title of https://example.com?').run_sync().final_result())

OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 uv run agent.py

To trace to Logfire, follow Pydantic AI’s Logfire guide (logfire.instrument_pydantic_ai()).

Spans carry the task, the code the model writes and what it reads from pages. Typed sensitive_data values are redacted from cell output, but screenshots and page text are not. Pass InstrumentationSettings(..., include_content=False) to keep message content out of the spans.

Not in 1.0

Laminar and OpenLIT integrations. 0.x shipped Laminar hooks and an OpenLIT guide. 1.0 has neither; use the OpenTelemetry spans above, which any OTLP backend accepts. The 0.x pages: observability (0.x), OpenLIT (0.x).
Per-step hooks and per-action history. 0.x on_step_start / on_step_end and history.action_names() are gone. Use the log handler and result.messages.

browser-use’s own anonymized usage events are a separate thing: see Telemetry.

Get Started

Core

Safety

Features

Integrations

Reference

Logs

Usage and cost

Transcript and workspace

OpenTelemetry traces

Not in 1.0

​Logs

​Usage and cost

​Transcript and workspace

​OpenTelemetry traces

​Not in 1.0

Logs

Usage and cost

Transcript and workspace

OpenTelemetry traces

Not in 1.0