---
title: "AI Agent Audit Trails and RBAC: Five Platforms Compared"
slug: "ai-agent-audit-trail-rbac-governance-platforms-compared"
author: "Miguel Amigot"
date: "2026-11-19 12:00:00"
category: "Premium"
sector: "cross"
topics: "AI governance platform, AI agent audit trail, role-based access control, RBAC for AI agents, agent identity, AWS Bedrock AgentCore, Microsoft Entra Agent ID, Google Cloud Agent Identity, LangSmith, self-hosted AI agents, AI agent governance"
summary: "Which AI governance platforms give AI agents an audit trail and role-based access control? A side-by-side read of the vendor documentation for AWS Bedrock AgentCore, Microsoft Entra Agent ID, Google Cloud Agent Identity, LangSmith and ibl.ai: what each one logs, which plan or setting turns it on, and whose infrastructure holds the record."
banner: ""
thumbnail: ""
linkedin: |
  "Does it have an audit trail and RBAC?" is the wrong question to ask an AI agent platform. Every serious one says yes. The useful questions are what gets logged, what turns it on, and whose infrastructure holds the record.

  We read the vendor documentation for five platforms. A few things it says:

  → AWS Bedrock AgentCore Gateway can record each tool call in CloudTrail, but data events are not captured by default and carry additional charges.
  → LangSmith audit logs and RBAC are Enterprise-plan features, with audit logs retained for up to 400 days.
  → Microsoft Entra Agent ID is available to all Entra customers, but extending Entra security features to agents requires Microsoft Agent 365, licensed for each user.

  Two layers matter: who changed the agent, and who may use, edit or watch it. On ibl.ai each agent has its own change log (who changed which setting, and when) and its own roles (Editor, Chat, Analytics Viewer), on top of organization-wide RBAC policies.

  With ibl.ai you own all the code and the data, so the audit log and the access policies run on infrastructure you control, on any model.

  #iblai #AgenticAI #EnterpriseAI #AIGovernance #RBAC #AuditTrail
faqs:
  - question: "Which AI governance platforms give AI agents an audit trail and role-based access control?"
    answer: "AWS Bedrock AgentCore, Microsoft Entra Agent ID, Google Cloud Agent Identity, LangSmith and ibl.ai all document some form of both. They differ in what is logged (configuration changes, authentication, tool calls), in what turns it on (a plan, a license or a setting), and in whose infrastructure stores the record."
  - question: "What should an AI agent audit trail record?"
    answer: "Two things: who changed the agent's configuration and when, and what the agent did while running. Access control should answer who may use, edit or review each agent, and what systems the agent itself may reach."
  - question: "Who holds the audit log when an AI agent platform is self-hosted?"
    answer: "You do. When the platform runs on infrastructure you operate, the audit log, the transcripts and the access policies are stored there too, so producing evidence for a regulator does not depend on a vendor's retention settings or export tools."
---

## The Short Answer

**Audit trails and role-based access control for AI agents are documented by AWS Bedrock AgentCore, Microsoft Entra Agent ID, Google Cloud Agent Identity, LangSmith and ibl.ai, but they differ in what gets logged, which plan or setting unlocks it and whose infrastructure holds the record. On ibl.ai you own all the code and the data, so agent audit logs and access policies stay inside your perimeter.**

Most governance listicles compare model and policy tools: risk registers, model cards, bias tests. This post answers a narrower question that buyers now type word for word: which platform gives the agents themselves an audit trail and RBAC?

Every vendor fact below comes from that vendor's own documentation, read on 18 November 2026. Where a capability is not on the pages we read, we say so rather than guess.

## What should an audit trail for AI agents record?

An audit trail for AI agents has two layers, and most products cover one better than the other.

The first is the change log: who altered the agent's model, prompt, tools or access, and when. The second is the action record: what the agent did while running, which tools it called and on whose behalf.

Access control also has two directions. Inward: which people may use, edit or review an agent. Outward: which systems the agent itself may reach. A platform that scores well on one direction can be thin on the other.

We set out the day-one version of these controls, including the agent registry that has to come first, in [Shadow IT Stored Data. Shadow Agents Take Actions.](/blog/agent-governance-new-shadow-it-day-one-controls) This post maps them to five named platforms.

## Which AI governance platforms give agents an audit trail and RBAC?

All five document some form of both. The table compresses what each vendor's pages say; the sections below give the detail and the source for every cell.

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Platform</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Access control for agents</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Audit trail described</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">What turns it on</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Where the record lives</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>AWS Bedrock AgentCore</strong></td>
      <td style="padding:0.75rem;">Workload identities for agents via AgentCore Identity</td>
      <td style="padding:0.75rem;">CloudTrail management events; Gateway tool calls as data events, arguments and results redacted</td>
      <td style="padding:0.75rem;">Data events off by default, additional charges</td>
      <td style="padding:0.75rem;">An S3 bucket you choose in your AWS account</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Microsoft Entra Agent ID</strong></td>
      <td style="padding:0.75rem;">Agent identities that can hold Azure RBAC, Entra directory and app roles</td>
      <td style="padding:0.75rem;">Authentication by Copilot Studio agents logged in Entra ID; creator recorded as sponsor</td>
      <td style="padding:0.75rem;">Agent ID for all Entra customers; Entra security features for agents need Agent 365, licensed per user</td>
      <td style="padding:0.75rem;">Your Microsoft Entra tenant</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Google Cloud Agent Identity</strong></td>
      <td style="padding:0.75rem;">A SPIFFE ID per agent; IAM allow and deny policies</td>
      <td style="padding:0.75rem;">Logs show both the agent's and the user's identity when acting on a user's behalf</td>
      <td style="padding:0.75rem;">You manage the IAM bindings, and re-grant them after a redeploy</td>
      <td style="padding:0.75rem;">Your Google Cloud project's audit logging</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>LangSmith</strong></td>
      <td style="padding:0.75rem;">Workspace roles (Admin, Editor, Viewer), custom roles and attribute-based access policies</td>
      <td style="padding:0.75rem;">Changes to organization settings, membership, credentials and workspaces; retained up to 400 days</td>
      <td style="padding:0.75rem;">RBAC and audit logs on Enterprise plans</td>
      <td style="padding:0.75rem;">LangSmith cloud, or your cluster if self-hosted</td>
    </tr>
    <tr style="background:#f0f9ff; border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>ibl.ai</strong></td>
      <td style="padding:0.75rem;">Per-agent roles (Editor, Chat, Analytics Viewer) plus organization RBAC roles, policies and groups</td>
      <td style="padding:0.75rem;">Per-agent change log of who changed which setting and when; conversation transcripts with user and model</td>
      <td style="padding:0.75rem;">Built in; the audit view is gated by a view-audit-logs permission</td>
      <td style="padding:0.75rem;">Self-hosted: infrastructure you control (your cloud, on-premise or air-gapped); hosted: ibl.ai's cloud</td>
    </tr>
  </tbody>
</table>

## How does AWS Bedrock AgentCore log and scope AI agents?

AWS describes [AgentCore Identity](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/identity.html) as identity and credential management built for AI agents. Agent identities are implemented as workload identities, so an agent authenticates as itself rather than borrowing a person's credentials.

The action record runs through CloudTrail. The [AgentCore Gateway CloudTrail guide](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/understanding-gateway-cloudtrail-log-entries.html) shows an `InvokeGateway` data event naming the tool called, with the request arguments and the response content redacted.

That record is not on by default. AWS states that Gateway [data event logging](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/enabling-cloudtrail-data-event-logging.html) must be explicitly enabled through a manually created trail backed by an S3 bucket, incurs additional charges, and can produce thousands of logs in a short time on a busy gateway.

The upside is placement: the logs land in an S3 bucket of your choosing, and AWS suggests a bucket in a separate account to centralize them. The gateway that writes them is an AWS managed service.

## How does Microsoft Entra Agent ID govern AI agent access?

Microsoft's [agent identities](https://learn.microsoft.com/en-us/entra/agent-id/what-are-agent-identities) are accounts in Microsoft Entra ID built for AI agents. An agent can act autonomously with rights given to its own identity, including Azure RBAC roles, Entra directory roles and app roles, or act on a user's behalf with delegated rights.

On auditability, Microsoft states that each Copilot Studio agent gets an agent identity, the user who created it is recorded as its sponsor, and any authentication these agents perform is logged in Entra ID as AI agents.

The licensing is where buyers should read closely. Agent ID is available for all Microsoft Entra customers, but Microsoft says extending Entra security features to agents requires Microsoft Agent 365, which requires a license for each user.

A per-user license is the wrong shape for agent governance. Agents multiply with workflows, not headcount, so the bill tracks the number of people while the risk tracks the number of agents.

## How does Google Cloud's Agent Identity handle AI agent permissions and logs?

Google's [Agent Identity](https://docs.cloud.google.com/agent-builder/agent-engine/agent-identity) assigns each agent a unique SPIFFE ID, and because the agent itself is the principal, permissions are granted directly to it. Identities are not shared by multiple workloads by default.

Access control runs through standard IAM allow and deny policies. Google notes that you must manage the IAM bindings yourself, and that after a redeploy the required roles have to be granted to the new principal, since old bindings do not carry over.

For the audit trail, Google says Agent Identity integrates with audit logging, and that when an agent acts on a user's behalf the logs show both identities. Which specific operations are recorded, and at what retention, is not described on the page we read.

## How does LangSmith handle audit logs and role-based access?

LangChain's LangSmith documents RBAC in its [administration overview](https://docs.langchain.com/langsmith/administration-overview), and states that RBAC is only available to Enterprise customers. Three built-in workspace roles exist: Workspace Admin, Workspace Editor and Workspace Viewer, and organization admins can create custom roles.

Its [audit logs](https://docs.langchain.com/langsmith/audit-logs) are also an Enterprise-plan feature. They record changes to organization settings, membership, credentials, workspaces and other resources, each with timestamp, actor, operation, affected resources and outcome.

Audit logs are retained for up to 400 days, returned in the OCSF API Activity format, and available on self-hosted LangSmith from Helm chart 0.12.33. Only Organization Admins and Organization Operators can read them; workspace roles give no access.

The scope is the account and its resources. LangSmith also lists attribute-based access policies for fine-grained permissions, which the pages we read do not detail. A per-agent change history or per-agent roles are not described there.

## How does ibl.ai give AI agents an audit trail and role-based access control?

On ibl.ai, every agent has its own [Audit panel](/docs/os/agent-settings/audit): a change log of who changed what and when. Each entry carries the user's email, a plain-language description such as "Updated LLM model, LLM provider" or "Enabled tools", the agent's unique ID and a timestamp.

The log filters by user, by date range and by action type (create, update or delete). It is admin-gated: the audit view appears only for people whose role grants the view-audit-logs permission on that agent.

Access is set per agent in the [Access panel](/docs/os/agent-settings/access). Three roles can be granted on each agent: **Editor** (update configuration, prompts and tools), **Chat** (conversations without editing) and **Analytics Viewer** (the agent's analytics). Roles go to named users, emails or whole groups.

Above the agent sits organization-wide RBAC. A role is a named bundle of permission strings with wildcard support and field-level data actions, and a [policy](/docs/os/organization-settings/policies) grants that role to users or groups on hierarchical resources, from every agent in the organization down to one.

For what the agent did in conversation, the transcripts record the user, the model, message count and estimated cost, with the full message thread. A separate log of individual tool calls is not described in the ibl.ai documentation we cite.

What an agent may reach is set in its [sandbox](/docs/os/agent-settings/sandbox), where a virtual machine starts with no network and a custom allowlist takes at most 100 exact host:port entries.

## Who holds the AI agent audit log when the platform is self-hosted?

You do, and that is the difference the table above cannot fully show.

On a managed platform the services that issue identities and write the log run on the vendor's infrastructure, so what gets logged depends on its settings, plans and formats, even when the log lands in your account.

ibl.ai runs in your cloud, on-premise or [fully air-gapped](/service/air-gapped-ai), on any model.

The change log, transcripts and RBAC policies are rows in a database you operate, so a regulator's request is a query you run, not a ticket you file. With the Codebase Transfer engagement, the source code is yours too.

That matters because agent logging standards are still moving. As we documented in [AI Agent Observability Has No Standard to Buy Yet](/blog/agent-observability-fragmenting-stack-own-the-audit-trail), the dedicated repository for OpenTelemetry's GenAI conventions had zero tagged releases as of October 2026, so owning the store is the safest bet.

ibl.ai is family-owned and operated from New York, and our team stays on after deployment for platform updates, new agents and ongoing services.

## What questions should a buyer ask an AI agent governance vendor?

Five questions separate a documented capability from one you can rely on in 2026:

1. **Does each agent have its own change log**, naming who changed its model, prompt or tools, and when?
2. **Can roles be granted per agent**, or only per workspace or account?
3. **Is action logging on by default**, or does it need a setting, a plan or an extra charge?
4. **Which plan or license unlocks RBAC and audit logs**, and is it priced per user?
5. **Where is the record stored**, and for how long, if the contract ends?

Ask each vendor to answer from its own documentation. Where the answer is "not described", treat it as unknown until the vendor shows you the page. To see how ibl.ai answers all five, start with [Agentic OS](/product/agentic-os).

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
