Skip to content
← Guides
Guide

How to find and inventory every AI agent in your company

Inventory AI agents by combining platform registries, identity and OAuth records, cloud and network logs, and owner attestation into one register.

By Sycamore.

Key takeaways

  • An AI agent inventory is one register of every agent the organization builds, buys or lets employees create, with an owner, purpose, models, tools, data access and identity for each.
  • No single console shows every agent. Each platform’s console sees only the agents built on or connected to that platform, so a complete inventory combines several sources.
  • Agents leave traces in four places, and a complete inventory checks all four: platform admin consoles, the identity provider, cloud and network logs, and the people who built or bought them.
  • Automated discovery produces candidates, not confirmed agents. Each candidate needs a named owner to confirm what it is and what it does.
  • An inventory stays accurate only if new agents must be registered before they get production access, and discovery keeps running to catch the ones that were not.

An AI agent inventory is a single register of every AI agent your organization runs, whether built in-house, bought or created by employees in no-code tools, with an owner, purpose, models, tools, data access and identity recorded for each. To build one, combine four sources, because no single console shows every agent: the admin lists in each AI platform, the identity provider and app-consent records, cloud and network logs, and attestation from the teams that build and buy software. Then require every new agent to be registered before it gets production access. The same sources, especially OAuth grants, network monitoring and procurement records, also surface the unapproved AI tools employees use, often called shadow AI.

Why do you need an AI agent inventory?

You cannot govern an agent you do not know exists. Every other control, from approving a use case to revoking access in an incident, starts with knowing which agents are running, who owns them and what they can reach.

Frameworks treat the inventory as a basic governance step. The NIST AI Risk Management Framework Playbook lists it under GOVERN 1.6: “Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities.” It suggests inventorying all models or systems where possible, or at least high-risk ones, and naming a person or team responsible for maintaining the inventory.1 In the US federal government it is mandatory: OMB memorandum M-25-21 requires each agency, with some exceptions, to inventory its AI use cases at least annually and publish a public version.2

Unknown AI use has a measurable cost. IBM’s Cost of a Data Breach Report 2026, based on breaches at 602 organizations, found that the share of security incidents involving shadow AI, where workers use unapproved AI, more than doubled to 43% from 20% the year before, and that these incidents had higher average breach costs than the year before, USD 5.39 million against USD 4.63 million. It also found that 68% of the breached organizations lacked AI governance to manage AI or detect shadow AI.3

Which systems belong in an AI agent inventory?

Count any software that uses an AI model to decide what to do and then acts through tools, data or other systems, under its own credentials or a user’s. What matters for the inventory is what an agent can reach and do, not how it was built.

That covers more than the agents your engineers write:

  • Custom agents built with model APIs and agent frameworks, running in your cloud accounts or data centers.
  • Agents built on platforms your organization licenses, such as Microsoft Copilot Studio, Google Gemini Enterprise or AWS’s agent services.
  • Agents employees create in no-code builders inside those platforms, often without a developer involved. Gemini Enterprise’s Workflow Builder lets employees and non-admin users build and deploy chat agents and connect them to Google, third-party and custom data sources, within the connectors and permissions an administrator provisions.4 Copilot Studio agents can call connectors to Microsoft and non-Microsoft services as tools.5
  • Agents inside software you buy, where a vendor’s agent acts on your data with access your organization granted.
  • Tool servers agents connect to, such as Model Context Protocol (MCP) servers. MCP servers expose tools, which the protocol defines as executable functions that AI applications can invoke to perform actions.6 They extend what any connected agent can do, so they belong in the register too. For the controls to put on those servers once you have found them, see how to secure and govern MCP servers that AI agents use.

Where do AI agents hide?

Each platform’s console sees only the agents built on or connected to that platform, so the places to look follow from where agents run.

Where agents run Typical examples Where to look first
Productivity and business suites Copilot Studio agents, Gemini Enterprise agents The platform’s admin center or agent registry
Cloud AI services Agents on AWS, Google Cloud or Azure Each cloud account’s agent registry or API, and its audit logs
Custom code Agents built on model APIs and open-source libraries The identity provider’s app registrations, service accounts, the CMDB
Third-party apps and SaaS vendors Vendor agents acting on your data OAuth grants and app-consent records
Employee devices Local agents and local MCP servers Endpoint and network monitoring

Even within one vendor, two consoles can give two different counts. Microsoft’s documentation explains that the Agent 365 agent registry in the Microsoft 365 admin center gives tenant-wide visibility across agent platforms, while its Power Platform inventory shows only agents built on Power Platform, and that “each number is correct within its own context.”7 Treat every console as one source among several.

Local tools are the hardest to see. The MCP documentation notes that local MCP servers typically serve a single client, while remote servers typically serve many,6 and the public MCP Registry does not support private servers.8 A server running on a developer’s laptop will not appear in any central list unless you look for it.

How do you find every AI agent?

Run these steps in order. The first two find most agents quickly. The later ones find the ones that were never registered anywhere.

  1. Export each platform’s own list. The major agent platforms include an admin view of their agents. The agent registry in the Microsoft 365 admin center shows agents available to the organization, including counts of agents without owners and unmanaged agents, and exports to CSV.9 Gemini Enterprise gives administrators a view of the agents used by the organization, including those from Google, third parties and internal teams.10 AWS Agent Registry can auto-detect agents on AgentCore runtimes and gateways across an organization.11 Run account-level listings in every account and region, not just the main one.
  2. Search the identity provider. Custom agents usually authenticate like any other application. Microsoft’s guidance for finding them in Entra ID suggests scanning service principals for signals such as permissions to AI services, frequent sign-ins with no user present, and display names containing words like agent, bot or copilot. It warns that this produces candidates, not confirmed agents, and that custom agents with generic permissions and no naming pattern can be invisible to such scans.12
  3. Review OAuth grants and app consents. Third-party tools, including vendor agents, often reach company data through OAuth. Google Workspace lets administrators review apps that have accessed Google data, a list Google says is updated 48 hours after a token is granted or revoked, and its OAuth log records each time a third-party app is authorized to access account data.1314 Microsoft Entra lets administrators review permissions granted to applications through user or admin consent.15
  4. Read cloud and network logs. Model and agent calls leave records in cloud audit logs, but check what is logged by default. In AWS CloudTrail, Amazon Bedrock model calls such as InvokeModel are logged as management events, while agent calls such as InvokeAgent are data events, which CloudTrail does not log by default.16 Turn on the logging you need before relying on it. Network and endpoint tools that classify AI traffic can find agents on devices that no registry knows about.
  5. Ask the people who build and buy software. Automated methods miss things. Microsoft’s own discovery guidance falls back on organizational sources such as the configuration management database and developer attestation.12 Ask engineering teams to declare the agents they run. Check procurement and expense records for AI tools bought outside the usual process, and add an AI question to vendor reviews.
  6. Reconcile into one register. Merge the results, remove duplicates (the same agent often appears in a platform console, the identity provider and the logs), and confirm each candidate with a named owner before recording it as an agent.

What should each inventory record contain?

Record enough to answer who owns the agent, what it is for, what it can reach and how to stop it. The NIST Playbook describes an AI system inventory as an organized database of artifacts that may include system documentation, incident response plans, data dictionaries, links to source code, and names and contact information for the people involved.1 For agents, a useful record has these fields:

  • Identity: a unique ID, name, and the credentials or service account it runs under.
  • Owner: a named person accountable for the agent, plus the team that operates it.
  • Purpose and use case: what it does, what decisions it makes, and its risk tier.
  • Source: built in-house, built on a platform, created by an employee, or supplied by a vendor.
  • Models: which models and versions it uses, and through which provider.
  • Tools and connections: the tools, APIs, MCP servers and other agents it can call.
  • Data access: which data sources it can read or write, and how sensitive they are.
  • Environments and status: pilot, production or retired, and where it runs.
  • Approvals and reviews: who approved it, when, and when it is next due for review.
  • Observed activity: when it last ran and what it actually called, compared with what it was approved to do.

That last field matters. A record of what an agent was approved to do is only useful if you can compare it with what the agent does in practice.

How do you keep the inventory current?

An inventory goes stale quickly, because agents are cheap to create. Keep it accurate with two habits that work together.

First, make registration a condition of access. New agents get production credentials, connectors and budgets only once they are in the register with an owner. This catches the agents built through approved routes, and fits the readiness checks in our guide on taking an AI agent from prototype to production. Joint guidance on agentic AI from cybersecurity agencies in Australia, Canada, New Zealand, the United Kingdom and the United States takes the same line: in its advice on agent identity management, it recommends denying access to “any agent or cryptographic key that is not present in the trusted registry.”17

Second, keep discovery running and compare its results with the register. The same guidance says to “periodically reconcile the registry against the live set of agents.”17 Any agent that discovery finds and the register does not is a finding to resolve: register it, restrict it or retire it. Unapproved AI tools that are not agents need the same treatment. Offer approved alternatives for common needs, use your identity provider’s user consent settings to restrict which apps users can approve on their own and route the rest to an administrator for review,18 and block tools that send sensitive data where it should not go.

Report agents without owners as their own category. Microsoft’s registry counts them separately, which is a sensible default.9

When an agent is retired, keep its record rather than deleting it. The NIST Playbook notes that decommissioned systems may be stored in an inventory alongside active ones for an established length of time,1 which keeps the history available for incidents and audits.

The inventory is the first of the controls in our guide to governing AI agents in production, which covers what to do once you know what is running: approvals, identity, runtime policy and stopping an agent.

Where Sycamore fits

Sycamore Govern is an AI governance control plane that brings the AI systems an organization builds or buys into one register, approves their use, enforces policy at runtime, and preserves the evidence. Its discovery step inventories AI use cases, agents, applications, models, connectors and MCP servers. Systems made in Sycamore Build register automatically, and external agents join through telemetry, the Agent2Agent (A2A) protocol or direct registration. Each record connects ownership, versions, capabilities, tools, data access and activity, and separates declared intent from observed behavior.

Teams can start with the registry and intake on their own and add runtime controls, cost management or audit evidence later.

Frequently asked questions

Is an AI agent inventory the same as an AI model inventory?

No. A model inventory lists the models an organization trains or uses. An agent inventory lists the systems that use those models to take actions, and records what each one can reach, such as tools, data and other systems, and under which identity. Most organizations need both, and each agent record should point to the models it uses.

Should agents employees build in no-code tools be in the inventory?

Yes, if they can act on company data or systems. No-code builders let employees connect agents to company data and services through connectors, within whatever limits administrators have set, and such an agent may act with the permissions of the person using it or of its creator. Record them, and set a rule for which ones need a full review based on what they can access.

How often should we run discovery?

Continuously where the platform supports it, and at least on a fixed schedule where it does not. Agents are cheap to create, so an inventory refreshed once a year or even once a quarter will miss some. Match the schedule to how each source updates, since some admin reports and logs update with a delay.

What about AI tools employees use that are not agents?

Find them with the same sources, especially OAuth grants, network monitoring, and procurement and expense records. Then decide per tool. Offer an approved alternative for common needs, set a consent policy so new apps need approval before they can reach company data, and block tools that send sensitive data where it should not go. Add a tool to the register once it can act on company systems or data, not only answer questions.

What should happen to an agent nobody will claim?

Pause or restrict it until someone accepts ownership. An agent without an owner has no one to answer for what it does, approve changes to it or respond if it misbehaves. Record the decision, and if the agent is retired, keep its inventory record and logs for the period your retention policy requires.

Sources

  1. NIST Trustworthy and Responsible AI Resource Center, “AI RMF Playbook: Govern”, https://airc.nist.gov/airmf-resources/playbook/govern/

  2. Office of Management and Budget, “M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust” (April 3, 2025), https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf

  3. IBM and Ponemon Institute, “Cost of a Data Breach Report 2026”, https://assets.ibm.com/adobe/assets/urn:aaid:aem:21111142-1251-4369-86fb-57b82f5bb108/original/as/Cost%20of%20a%20Data%20Breach%20Report%202026.pdf

  4. Google Cloud, “Workflow Builder overview”, Gemini Enterprise documentation, https://docs.cloud.google.com/gemini/enterprise/docs/workflow-builder

  5. Microsoft Learn, “Use connectors in Copilot Studio agents”, https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-connectors

  6. Model Context Protocol, “Architecture overview”, https://modelcontextprotocol.io/docs/learn/architecture

  7. Microsoft Learn, “Understand agent counts across admin surfaces”, https://learn.microsoft.com/en-us/power-platform/admin/inventory-agent-counts

  8. Model Context Protocol, “The MCP Registry”, https://modelcontextprotocol.io/registry/about

  9. Microsoft Learn, “Agent Registry in Microsoft 365 admin center”, https://learn.microsoft.com/en-us/microsoft-agent-365/admin/agent-registry-overview

  10. Google Cloud, “Agents overview”, Gemini Enterprise documentation, https://docs.cloud.google.com/gemini/enterprise/docs/agents-overview

  11. AWS, “AWS Agent Registry is now generally available” (August 31, 2026), https://aws.amazon.com/about-aws/whats-new/2026/08/aws-agent-registry-generally-available/

  12. Microsoft Learn, “Migrate custom app registrations to Agent ID”, https://learn.microsoft.com/en-us/entra/agent-id/migrate-custom-app-registrations-to-agent-id

  13. Google Workspace Admin Help, “Control which third-party & internal apps access Google Workspace data”, https://knowledge.workspace.google.com/admin/apps/control-which-apps-access-google-workspace-data

  14. Google Workspace Admin Help, “OAuth log events”, https://knowledge.workspace.google.com/admin/reports/oauth-log-events

  15. AWS, “Monitor Amazon Bedrock API calls using CloudTrail”, Amazon Bedrock User Guide, https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html

  16. Australian Signals Directorate’s Australian Cyber Security Centre, CISA, NSA, Canadian Centre for Cyber Security, NCSC-NZ and NCSC-UK, “Careful adoption of agentic AI services” (2026), https://www.ncsc.govt.nz/assets/guidance/Documents/Careful-adoption-of-agentic-AI-services_FINAL.pdf

# How to find and inventory every AI agent in your company

Inventory AI agents by combining platform registries, identity and OAuth records, cloud and network logs, and owner attestation into one register.

## Key takeaways

- An AI agent inventory is one register of every agent the organization builds, buys or lets employees create, with an owner, purpose, models, tools, data access and identity for each.
- No single console shows every agent. Each platform's console sees only the agents built on or connected to that platform, so a complete inventory combines several sources.
- Agents leave traces in four places, and a complete inventory checks all four: platform admin consoles, the identity provider, cloud and network logs, and the people who built or bought them.
- Automated discovery produces candidates, not confirmed agents. Each candidate needs a named owner to confirm what it is and what it does.
- An inventory stays accurate only if new agents must be registered before they get production access, and discovery keeps running to catch the ones that were not.

An AI agent inventory is a single register of every AI agent your organization runs, whether built in-house, bought or created by employees in no-code tools, with an owner, purpose, models, tools, data access and identity recorded for each. To build one, combine four sources, because no single console shows every agent: the admin lists in each AI platform, the identity provider and app-consent records, cloud and network logs, and attestation from the teams that build and buy software. Then require every new agent to be registered before it gets production access. The same sources, especially OAuth grants, network monitoring and procurement records, also surface the unapproved AI tools employees use, often called shadow AI.

## Why do you need an AI agent inventory?

You cannot govern an agent you do not know exists. Every other control, from approving a use case to revoking access in an incident, starts with knowing which agents are running, who owns them and what they can reach.

Frameworks treat the inventory as a basic governance step. The NIST AI Risk Management Framework Playbook lists it under GOVERN 1.6: "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities." It suggests inventorying all models or systems where possible, or at least high-risk ones, and naming a person or team responsible for maintaining the inventory.[^nist-playbook] In the US federal government it is mandatory: OMB memorandum M-25-21 requires each agency, with some exceptions, to inventory its AI use cases at least annually and publish a public version.[^omb]

Unknown AI use has a measurable cost. IBM's Cost of a Data Breach Report 2026, based on breaches at 602 organizations, found that the share of security incidents involving shadow AI, where workers use unapproved AI, more than doubled to 43% from 20% the year before, and that these incidents had higher average breach costs than the year before, USD 5.39 million against USD 4.63 million. It also found that 68% of the breached organizations lacked AI governance to manage AI or detect shadow AI.[^ibm-2026]

## Which systems belong in an AI agent inventory?

Count any software that uses an AI model to decide what to do and then acts through tools, data or other systems, under its own credentials or a user's. What matters for the inventory is what an agent can reach and do, not how it was built.

That covers more than the agents your engineers write:

- **Custom agents** built with model APIs and agent frameworks, running in your cloud accounts or data centers.
- **Agents built on platforms** your organization licenses, such as Microsoft Copilot Studio, Google Gemini Enterprise or AWS's agent services.
- **Agents employees create** in no-code builders inside those platforms, often without a developer involved. Gemini Enterprise's Workflow Builder lets employees and non-admin users build and deploy chat agents and connect them to Google, third-party and custom data sources, within the connectors and permissions an administrator provisions.[^gemini-wb] Copilot Studio agents can call connectors to Microsoft and non-Microsoft services as tools.[^copilot-connectors]
- **Agents inside software you buy**, where a vendor's agent acts on your data with access your organization granted.
- **Tool servers agents connect to**, such as Model Context Protocol (MCP) servers. MCP servers expose tools, which the protocol defines as executable functions that AI applications can invoke to perform actions.[^mcp-arch] They extend what any connected agent can do, so they belong in the register too. For the controls to put on those servers once you have found them, see [how to secure and govern MCP servers that AI agents use](/guides/secure-mcp-servers-for-agents/).

## Where do AI agents hide?

Each platform's console sees only the agents built on or connected to that platform, so the places to look follow from where agents run.

| Where agents run                  | Typical examples                                     | Where to look first                                                   |
| --------------------------------- | ---------------------------------------------------- | --------------------------------------------------------------------- |
| Productivity and business suites  | Copilot Studio agents, Gemini Enterprise agents      | The platform's admin center or agent registry                         |
| Cloud AI services                 | Agents on AWS, Google Cloud or Azure                 | Each cloud account's agent registry or API, and its audit logs        |
| Custom code                       | Agents built on model APIs and open-source libraries | The identity provider's app registrations, service accounts, the CMDB |
| Third-party apps and SaaS vendors | Vendor agents acting on your data                    | OAuth grants and app-consent records                                  |
| Employee devices                  | Local agents and local MCP servers                   | Endpoint and network monitoring                                       |

Even within one vendor, two consoles can give two different counts. Microsoft's documentation explains that the Agent 365 agent registry in the Microsoft 365 admin center gives tenant-wide visibility across agent platforms, while its Power Platform inventory shows only agents built on Power Platform, and that "each number is correct within its own context."[^ms-counts] Treat every console as one source among several.

Local tools are the hardest to see. The MCP documentation notes that local MCP servers typically serve a single client, while remote servers typically serve many,[^mcp-arch] and the public MCP Registry does not support private servers.[^mcp-registry] A server running on a developer's laptop will not appear in any central list unless you look for it.

## How do you find every AI agent?

Run these steps in order. The first two find most agents quickly. The later ones find the ones that were never registered anywhere.

1. **Export each platform's own list.** The major agent platforms include an admin view of their agents. The agent registry in the Microsoft 365 admin center shows agents available to the organization, including counts of agents without owners and unmanaged agents, and exports to CSV.[^ms-registry] Gemini Enterprise gives administrators a view of the agents used by the organization, including those from Google, third parties and internal teams.[^gemini] AWS Agent Registry can auto-detect agents on AgentCore runtimes and gateways across an organization.[^aws-registry] Run account-level listings in every account and region, not just the main one.
2. **Search the identity provider.** Custom agents usually authenticate like any other application. Microsoft's guidance for finding them in Entra ID suggests scanning service principals for signals such as permissions to AI services, frequent sign-ins with no user present, and display names containing words like agent, bot or copilot. It warns that this produces candidates, not confirmed agents, and that custom agents with generic permissions and no naming pattern can be invisible to such scans.[^entra-migrate]
3. **Review OAuth grants and app consents.** Third-party tools, including vendor agents, often reach company data through OAuth. Google Workspace lets administrators review apps that have accessed Google data, a list Google says is updated 48 hours after a token is granted or revoked, and its OAuth log records each time a third-party app is authorized to access account data.[^google-apps][^google-oauth] Microsoft Entra lets administrators review permissions granted to applications through user or admin consent.[^entra-consent]
4. **Read cloud and network logs.** Model and agent calls leave records in cloud audit logs, but check what is logged by default. In AWS CloudTrail, Amazon Bedrock model calls such as InvokeModel are logged as management events, while agent calls such as InvokeAgent are data events, which CloudTrail does not log by default.[^cloudtrail] Turn on the logging you need before relying on it. Network and endpoint tools that classify AI traffic can find agents on devices that no registry knows about.
5. **Ask the people who build and buy software.** Automated methods miss things. Microsoft's own discovery guidance falls back on organizational sources such as the configuration management database and developer attestation.[^entra-migrate] Ask engineering teams to declare the agents they run. Check procurement and expense records for AI tools bought outside the usual process, and add an AI question to vendor reviews.
6. **Reconcile into one register.** Merge the results, remove duplicates (the same agent often appears in a platform console, the identity provider and the logs), and confirm each candidate with a named owner before recording it as an agent.

## What should each inventory record contain?

Record enough to answer who owns the agent, what it is for, what it can reach and how to stop it. The NIST Playbook describes an AI system inventory as an organized database of artifacts that may include system documentation, incident response plans, data dictionaries, links to source code, and names and contact information for the people involved.[^nist-playbook] For agents, a useful record has these fields:

- **Identity:** a unique ID, name, and the credentials or service account it runs under.
- **Owner:** a named person accountable for the agent, plus the team that operates it.
- **Purpose and use case:** what it does, what decisions it makes, and its risk tier.
- **Source:** built in-house, built on a platform, created by an employee, or supplied by a vendor.
- **Models:** which models and versions it uses, and through which provider.
- **Tools and connections:** the tools, APIs, MCP servers and other agents it can call.
- **Data access:** which data sources it can read or write, and how sensitive they are.
- **Environments and status:** pilot, production or retired, and where it runs.
- **Approvals and reviews:** who approved it, when, and when it is next due for review.
- **Observed activity:** when it last ran and what it actually called, compared with what it was approved to do.

That last field matters. A record of what an agent was approved to do is only useful if you can compare it with what the agent does in practice.

## How do you keep the inventory current?

An inventory goes stale quickly, because agents are cheap to create. Keep it accurate with two habits that work together.

First, make registration a condition of access. New agents get production credentials, connectors and budgets only once they are in the register with an owner. This catches the agents built through approved routes, and fits the readiness checks in our guide on [taking an AI agent from prototype to production](/guides/ai-agents-prototype-to-production/). Joint guidance on agentic AI from cybersecurity agencies in Australia, Canada, New Zealand, the United Kingdom and the United States takes the same line: in its advice on agent identity management, it recommends denying access to "any agent or cryptographic key that is not present in the trusted registry."[^five-eyes]

Second, keep discovery running and compare its results with the register. The same guidance says to "periodically reconcile the registry against the live set of agents."[^five-eyes] Any agent that discovery finds and the register does not is a finding to resolve: register it, restrict it or retire it. Unapproved AI tools that are not agents need the same treatment. Offer approved alternatives for common needs, use your identity provider's user consent settings to restrict which apps users can approve on their own and route the rest to an administrator for review,[^entra-user-consent] and block tools that send sensitive data where it should not go.

Report agents without owners as their own category. Microsoft's registry counts them separately, which is a sensible default.[^ms-registry]

When an agent is retired, keep its record rather than deleting it. The NIST Playbook notes that decommissioned systems may be stored in an inventory alongside active ones for an established length of time,[^nist-playbook] which keeps the history available for incidents and audits.

The inventory is the first of the controls in our guide to [governing AI agents in production](/guides/how-to-govern-ai-agents/), which covers what to do once you know what is running: approvals, identity, runtime policy and stopping an agent.

## Where Sycamore fits

[Sycamore Govern](/govern/) is an AI governance control plane that brings the AI systems an organization builds or buys into one register, approves their use, enforces policy at runtime, and preserves the evidence. Its discovery step inventories AI use cases, agents, applications, models, connectors and MCP servers. Systems made in [Sycamore Build](/build/) register automatically, and external agents join through telemetry, the Agent2Agent (A2A) protocol or direct registration. Each record connects ownership, versions, capabilities, tools, data access and activity, and separates declared intent from observed behavior.

Teams can start with the registry and intake on their own and add runtime controls, cost management or audit evidence later.

[^nist-playbook]: NIST Trustworthy and Responsible AI Resource Center, "AI RMF Playbook: Govern", https://airc.nist.gov/airmf-resources/playbook/govern/

[^omb]: Office of Management and Budget, "M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust" (April 3, 2025), https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf

[^ibm-2026]: IBM and Ponemon Institute, "Cost of a Data Breach Report 2026", https://assets.ibm.com/adobe/assets/urn:aaid:aem:21111142-1251-4369-86fb-57b82f5bb108/original/as/Cost%20of%20a%20Data%20Breach%20Report%202026.pdf

[^mcp-arch]: Model Context Protocol, "Architecture overview", https://modelcontextprotocol.io/docs/learn/architecture

[^mcp-registry]: Model Context Protocol, "The MCP Registry", https://modelcontextprotocol.io/registry/about

[^ms-counts]: Microsoft Learn, "Understand agent counts across admin surfaces", https://learn.microsoft.com/en-us/power-platform/admin/inventory-agent-counts

[^ms-registry]: Microsoft Learn, "Agent Registry in Microsoft 365 admin center", https://learn.microsoft.com/en-us/microsoft-agent-365/admin/agent-registry-overview

[^gemini]: Google Cloud, "Agents overview", Gemini Enterprise documentation, https://docs.cloud.google.com/gemini/enterprise/docs/agents-overview

[^gemini-wb]: Google Cloud, "Workflow Builder overview", Gemini Enterprise documentation, https://docs.cloud.google.com/gemini/enterprise/docs/workflow-builder

[^copilot-connectors]: Microsoft Learn, "Use connectors in Copilot Studio agents", https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-connectors

[^aws-registry]: AWS, "AWS Agent Registry is now generally available" (August 31, 2026), https://aws.amazon.com/about-aws/whats-new/2026/08/aws-agent-registry-generally-available/

[^entra-migrate]: Microsoft Learn, "Migrate custom app registrations to Agent ID", https://learn.microsoft.com/en-us/entra/agent-id/migrate-custom-app-registrations-to-agent-id

[^entra-user-consent]: Microsoft Learn, "Configure how users consent to applications", https://learn.microsoft.com/en-us/entra/identity/enterprise-apps/configure-user-consent

[^google-apps]: Google Workspace Admin Help, "Control which third-party & internal apps access Google Workspace data", https://knowledge.workspace.google.com/admin/apps/control-which-apps-access-google-workspace-data

[^google-oauth]: Google Workspace Admin Help, "OAuth log events", https://knowledge.workspace.google.com/admin/reports/oauth-log-events

[^entra-consent]: Microsoft Learn, "Review permissions granted to enterprise applications", https://learn.microsoft.com/en-us/entra/identity/enterprise-apps/manage-application-permissions

[^five-eyes]: Australian Signals Directorate's Australian Cyber Security Centre, CISA, NSA, Canadian Centre for Cyber Security, NCSC-NZ and NCSC-UK, "Careful adoption of agentic AI services" (2026), https://www.ncsc.govt.nz/assets/guidance/Documents/Careful-adoption-of-agentic-AI-services_FINAL.pdf

[^cloudtrail]: AWS, "Monitor Amazon Bedrock API calls using CloudTrail", Amazon Bedrock User Guide, https://docs.aws.amazon.com/bedrock/latest/userguide/logging-using-cloudtrail.html

## Frequently asked questions

### Is an AI agent inventory the same as an AI model inventory?

No. A model inventory lists the models an organization trains or uses. An agent inventory lists the systems that use those models to take actions, and records what each one can reach, such as tools, data and other systems, and under which identity. Most organizations need both, and each agent record should point to the models it uses.

### Should agents employees build in no-code tools be in the inventory?

Yes, if they can act on company data or systems. No-code builders let employees connect agents to company data and services through connectors, within whatever limits administrators have set, and such an agent may act with the permissions of the person using it or of its creator. Record them, and set a rule for which ones need a full review based on what they can access.

### How often should we run discovery?

Continuously where the platform supports it, and at least on a fixed schedule where it does not. Agents are cheap to create, so an inventory refreshed once a year or even once a quarter will miss some. Match the schedule to how each source updates, since some admin reports and logs update with a delay.

### What about AI tools employees use that are not agents?

Find them with the same sources, especially OAuth grants, network monitoring, and procurement and expense records. Then decide per tool. Offer an approved alternative for common needs, set a consent policy so new apps need approval before they can reach company data, and block tools that send sensitive data where it should not go. Add a tool to the register once it can act on company systems or data, not only answer questions.

### What should happen to an agent nobody will claim?

Pause or restrict it until someone accepts ownership. An agent without an owner has no one to answer for what it does, approve changes to it or respond if it misbehaves. Record the decision, and if the agent is retired, keep its inventory record and logs for the period your retention policy requires.

Published November 12, 2026, updated November 15, 2026 by Sycamore.