Logging an agent

To work with Cloud Logging in agents when they are deployed, use one of the following methods:

• stdout / stderr: by default (without any additional set up), logs written to stdout and stderr will be routed to the log IDs reasoning_engine_stdout and reasoning_engine_stderr respectively. The limitation is that they have to be text.
• Python logging: the built-in Python logger can be integrated with Cloud Logging. Compared to writing to stdout or stderr, this supports structured logs and requires minimal set up.
• Cloud Logging client: users can write structured log, and has full control over the logger (e.g., setting the logName and resource type).

Write logs for an agent

When writing logs for an agent, determine the:

• severity: E.g. info, warn, error
• payload: the contents of the log (e.g. text or JSON)
• additional fields: for correlating across logs (e.g. trace/span, tags, labels)

For example, to log the input of each query when developing an agent:

stdout or stderr

from typing import Dict

class MyAgent:

 def set_up(self):
 # No set up required. The logs from stdout and stderr are routed to
 # `reasoning_engine_stdout` and `reasoning_engine_stderr` respectively.
 pass

 def query(self, input: Dict):
 import sys

 print(
 f"input: {input}",
 file=sys.stdout, # or sys.stderr
 )

Python Logging

from typing import Dict

class MyAgent:

 def set_up(self):
 import os
 import google.cloud.logging

 self.logging_client = google.cloud.logging.Client(project="PROJECT_ID")
 self.logging_client.setup_logging(
 name="LOG_ID", # the ID of the logName in Cloud Logging.
 resource=google.cloud.logging.Resource(
 type="aiplatform.googleapis.com/ReasoningEngine",
 labels={
 "location": "LOCATION",
 "resource_container": "PROJECT_ID",
 "reasoning_engine_id": os.environ.get("GOOGLE_CLOUD_AGENT_ENGINE_ID", ""),
 },
 ),
 )

 def query(self, input: Dict):
 import logging
 import json

 logging_extras = {
 "labels": {"foo": "bar"},
 "trace": "TRACE_ID",
 }

 logging.info( # or .warning(), .error()
 json.dumps(input),
 extra=logging_extras,
 )

Cloud Logging client

from typing import Dict

class MyAgent:

 def set_up(self):
 import os
 import google.cloud.logging

 self.logging_client = google.cloud.logging.Client(project="PROJECT_ID")
 self.logger = self.logging_client.logger(
 name="LOG_ID", # the ID of the logName in Cloud Logging.
 resource=google.cloud.logging.Resource(
 type="aiplatform.googleapis.com/ReasoningEngine",
 labels={
 "location": "LOCATION",
 "resource_container": "PROJECT_ID",
 "reasoning_engine_id": os.environ.get("GOOGLE_CLOUD_AGENT_ENGINE_ID", ""),
 },
 ),
 )

 def query(self, input: Dict):
 logging_extras = {
 "labels": {"foo": "bar"},
 "trace": "TRACE_ID",
 }

 self.logger.log_struct(
 input,
 severity="INFO", # or "DEBUG", "WARNING", "ERROR", "CRITICAL"
 **logging_extras,
 )

When the agent is deployed and queried, it will generate log entries. For example, the code

remote_agent = agent_engines.create(
 MyAgent(),
 requirements=["cloudpickle==3", "google-cloud-logging"],
)

remote_agent.query(input={"hello": "world"})

will generate a log entry similar to the following:

stdout or stderr

{
 "insertId": "67a3bb3b000cc2df444361ab",
 "textPayload": "input: {'hello': 'world'}",
 "resource": {
 "type": "aiplatform.googleapis.com/ReasoningEngine",
 "labels": {
 "location": "LOCATION",
 "resource_container": "PROJECT_ID",
 "reasoning_engine_id": "RESOURCE_ID"
 }
 },
 "timestamp": "2025-02-05T19:25:47.836319Z",
 "logName": "projects/PROJECT_ID/logs/aiplatform.googleapis.com%2Freasoning_engine_stdout", # or `*_stderr`
 "receiveTimestamp": "2025-02-05T19:25:47.842550772Z"
}

Python Logging

{
 "insertId": "1ek9a2jfqh777z",
 "jsonPayload": {"hello": "world"},
 "resource": {
 "type": "aiplatform.googleapis.com/ReasoningEngine",
 "labels": {
 "location": "LOCATION",
 "resource_container": "PROJECT_ID",
 "reasoning_engine_id": "RESOURCE_ID",
 }
 },
 "timestamp": "2025-02-05T20:30:19.348067Z",
 "severity": "INFO",
 "labels": {
 "foo": "bar",
 "python_logger": "root",
 },
 "logName": "projects/PROJECT_ID/logs/LOG_ID",
 "trace": "TRACE_ID",
 "receiveTimestamp": "2025-01-30T21:38:50.776813191Z"
}

Cloud Logging client

{
 "insertId": "1ek9a2jfqh777z",
 "jsonPayload": {"hello": "world"},
 "resource": {
 "type": "aiplatform.googleapis.com/ReasoningEngine",
 "labels": {
 "location": "LOCATION",
 "resource_container": "PROJECT_ID",
 "reasoning_engine_id": "RESOURCE_ID",
 }
 },
 "timestamp": "2025-01-30T21:38:50.776813191Z",
 "severity": "INFO",
 "labels": {"foo": "bar"},
 "logName": "projects/PROJECT_ID/logs/LOG_ID",
 "trace": "TRACE_ID",
 "receiveTimestamp": "2025-01-30T21:38:50.776813191Z"
}

View logs for an agent

You can view your log entries using the Logs Explorer:

1. To get permission to view logs in Logs Explorer, ask your administrator to grant you the Logs Viewer role (roles/logging.viewer) on your project.
2. Go to Logs Explorer in the Google Cloud console:

Go to Logs Explorer 3. Select your Google Cloud project (corresponding to PROJECT_ID) at the top of the page. 4. In Resource Type, select Vertex AI Reasoning Engine.

Building queries

You can use the Logs Explorer to build queries incrementally. Queries are commonly built based on the following considerations:

• timeline: to search for relevant log entries based on time
• scope: to search for relevant log entries based on canonical attributes
• resource: separate it from other types of resources in your project.
• type: shows up as "Vertex AI Reasoning Engine" in Logs Explorer and "aiplatform.googleapis.com/ReasoningEngine" in the log entry.
• labels: for the location (LOCATION), project PROJECT_ID and resource RESOURCE_ID.
• logName: The log to which the log entry belongs:
• The log entries at build-time have log ID reasoning_engine_build.
• The log entries for stdout and stderr have log ID reasoning_engine_stdout and reasoning_engine_stderr respectively.
• The log entries from python logging or Cloud Logging client will have custom log IDs based on your code in Write logs for an agent.
• trace and span: for the logs when tracing queries.
• severity: for the severity of the log entry.
• insertId: the unique identifier for a log entry.
• labels: A map of key, value pairs that provides additional information about the log entry. The labels can be user-defined or system-defined, and are useful for categorizing logs and make it easier to search for them in Logs Explorer.
• payload: the contents of the log entry.

The following is an example of a query for all INFO logs from a deployed agent with RESOURCE_ID:

resource.labels.reasoning_engine_id=RESOURCE_ID AND
severity=INFO

You can view it in Logs Explorer at

https://console.cloud.google.com/logs/query;query=severity%3DINFO%0Aresource.labels.reasoning_engine_id%3D%22RESOURCE_ID%22;duration=DURATION?project=PROJECT_ID

where the query has been appropriately url-encoded and the other parameters are as follows:

• DURATION: for example PT30M for the past 30 minutes (or PT10M for the past 10 minutes), and
• PROJECT_ID: the Google Cloud project.

For details, visit Build and save queries by using the Logging query language.

Note: The Logs Explorer doesn't support aggregate operations, like counting the number of log entries that contain a specific pattern. For more advanced queries that covers aggregate operations and more, visit Query logs for an agent.

Query logs for an agent

For a programmatic approach to query logs, there are two common options:

• Structured Query Language (SQL). Log Analytics lets you query log views or analytics views.
• Log views have a fixed schema which corresponds to log entries.
• Analytics views have a schema that is based on the results of a SQL query.
• Python. Call the Cloud Logging API through the client library for your programming language (Python in this case).

Python

from google.cloud import logging

logging_client = logging.Client(project="PROJECT_ID")
logger = logging_client.logger("LOG_ID") # E.g. "logging_client"
print("Listing entries for logger {}:".format(logger.name))
for entry in logger.list_entries(
 filter_="resource.labels.reasoning_engine_id=RESOURCE_ID" # Optional
):
 timestamp = entry.timestamp.isoformat()
 print("* {}: {}".format(timestamp, entry.payload))

Each entry will correspond to a LogEntry. For details on the input arguments to logger.list_entries, visit the API reference.

SQL

Log view:

SELECT *
FROM `PROJECT_ID.LOCATION.BUCKET_ID.LOG_VIEW_ID`

Analytics view:

SELECT *
FROM `analytics_view.PROJECT_ID.LOCATION.ANALYTICS_VIEW_ID`