Skip to content
AI Agent Access Control

The logs that show what an agent did are mostly off until someone turns them on, and they must name the agent

After an incident the question is which agent did what, to which data. This guide covers the three log layers on AWS, Google Cloud and Azure (control plane, data access, agent trace), which ones are off by default, and whether each entry names the agent.

See the ranking ยท All guides

Three layers of log answer three different questions

A control-plane log shows who changed a role, a setting or a policy. A data-level log shows which objects, rows or keys the agent read and wrote. An agent trace shows the steps the agent took: the model calls, the tool calls and the order. A review needs all three, because each answers something the others cannot. The control-plane log will not show that the agent read a customer table, and the trace will not show that someone widened the agent's role at midnight. Each layer also has a different default, and the defaults are where the gaps sit. The rest of this guide goes cloud by cloud.

AWS records API calls, but data events and model logs need switching on

AWS documents that CloudTrail captures all Amazon Bedrock API calls as events, including calls from the console and from code. For Agents Classic, InvokeAgent and InvokeInlineAgent calls are logged as data events that you must enable with advanced event selectors for the AWS::Bedrock::AgentAlias resource type. AWS documents that CloudTrail trails do not log data events by default, and that extra charges apply. Model invocation logging, which collects the full request and response data for supported calls, is disabled by default and can deliver to CloudWatch Logs or Amazon S3.

On AgentCore, metrics, spans and logs are stored in Amazon CloudWatch. AWS documents CloudTrail coverage for only some AgentCore capabilities, and coverage for Runtime API calls was none found on 2026-11-10 in the pages read. For a Lambda agent, the CloudTrail userIdentity element records the type of IAM identity that made each request. AWS documents that Runtime and other AgentCore data goes to CloudWatch, so retention and who can read those log groups are settings to check as well.

Google Cloud separates always-on admin logs from data access logs

Google documents Admin Activity audit logs as written for user-driven API calls that change configuration, and as impossible to disable. Data Access audit logs are disabled by default for all services except a few, so reads and writes of the agent's data appear only after you enable them. For Agent Runtime, Google documents that the agent's stdout and stderr go to Cloud Logging under the log IDs reasoning_engine_stdout and reasoning_engine_stderr, that agent identity shows up in logs across Google Cloud services, and that for user-delegated flows the logs show both the user and the agent. Cloud Logging is not supported for child resources such as Sessions and Memory Bank. Confirm the Data Access setting for each service the agent touches before the review.

Azure splits identity logs, activity logs and traces

Microsoft documents three sources for Foundry Agent Service: Entra sign-in logs, which show which agent identity obtained a token for which resource, including on behalf of which user; Entra audit logs, which record blueprint and agent identity creation and deletion; and Foundry tracing in the connected Application Insights resource for tool calls. The Azure Monitor activity log records management operations on Azure resources. Resource logs, which carry data operations such as blob reads, are not collected until you create a diagnostic setting. For Container Apps and AKS, console and system logs go to Log Analytics.

Cost and volume are the usual reasons logs stay off

Data-level logs are large, so they are the first thing switched off. AWS documents that CloudTrail data events are often high-volume, that trails do not log them by default, and that additional charges apply. AWS also documents that advanced event selectors can filter on fields such as eventName, so the log can be limited to the agent alias or the bucket that matters instead of the whole account. Google documents that Data Access audit logs can be configured per service and per project, folder or organization. The review should accept a narrow data log on the agent's own resources and refuse an empty one.

Where the logs go matters as much as whether they exist. AWS documents organization trails, which log events for all accounts in an organization. Google documents that entries written by Cloud Audit Logs are immutable. A log destination that the agent's own role can edit or delete cannot serve as evidence of what the agent did. Ask who administers the log destination, and whether that group includes anyone who can also change the agent's role.

Four checks make the logs usable in an investigation

Turn on organization-wide control-plane logging (check). Log reads and writes to the data the agent touches (check). Make each logged action traceable to one agent identity, which needs the identity work in the first guide to be done (check). Alert on denied calls and on changes to the agent's role or policies (check). Also decide where the logs live: a log the agent's own role can edit or delete is weak evidence. The map lists the logging entry for each platform.

Sources

  1. AWS: monitor Amazon Bedrock API calls using CloudTrail, read 10 Nov 2026
  2. AWS: Bedrock model invocation logging, read 10 Nov 2026
  3. AWS: logging data events in CloudTrail, read 10 Nov 2026
  4. Google Cloud: enable Data Access audit logs, read 10 Nov 2026
  5. Google Cloud: logging for Agent Runtime, read 10 Nov 2026
  6. Microsoft: agent identity in Foundry Agent Service, read 10 Nov 2026
  7. Microsoft: monitor Azure Blob Storage, read 10 Nov 2026

Last reviewed 16 Nov 2026.