---
title: What fields belong in an AI agent inventory?
url: https://opencontroller.lyzr.com/blog/ai-agent-inventory-fields
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What fields belong in an AI agent inventory?

**An AI agent inventory needs a minimum of about a dozen fields: a stable ID, name, platform, owner, purpose, status, version, model, tools and MCP servers, credentials, data it can reach and risk tier. An extended set adds evals, guardrails, cost, framework mappings and review dates. Fill as many fields as possible from automated discovery, and ask owners only for what a scanner cannot know.**

Most inventories fail in one of two ways. Either they have too few fields to answer an incident question ("what can this agent touch?"), or they have so many owner-supplied fields that nobody keeps them current. The answer is to split fields by where the data comes from.

## What is the minimum field set for an AI agent inventory?

These are the fields you need before an agent should be allowed near production. Each one answers a question you will be asked during an incident, an audit or a cost review.

| Field | Source | Why it matters |
|---|---|---|
| Agent ID | System-generated | A stable key that survives renames and platform moves |
| Name and description | Discovery, then owner | Lets people recognise the agent |
| Platform and location | Discovery | Cloud account, region, SaaS tenant or device where it runs |
| Owner (person) | Identity provider | Someone accountable; flags leavers automatically |
| Owning team | Identity provider | Escalation path when the owner is away |
| Business purpose | Owner | Lets reviewers judge whether access is proportionate |
| Status | Registry | Draft, pilot, production, quarantined, retired |
| Current version | Discovery or CI | Ties behaviour and evidence to a specific build |
| Models used | Discovery or gateway | Provider, model name and endpoint |
| Tools and MCP servers | Discovery | What actions the agent can take |
| Credentials and identity | Discovery | API keys, service accounts, OAuth clients, workload identities |
| Data access | Discovery plus owner | Systems and data classes the agent can read or write |
| Risk tier | Policy, then reviewer | Drives which controls and reviews apply |

If you can only keep thirteen fields current, keep these.

## What does the extended field set add?

Add these once the minimum set is reliable. They turn the inventory from a list into evidence.

1. **User population.** Internal staff, customers, partners or other agents. Customer-facing agents usually need stricter controls.
2. **Acts on behalf of users.** Yes or no, and how delegation works. See [AI agent identity and on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation).
3. **Human approval points.** Which actions require a person to approve, and who.
4. **Eval history.** Last test date, test suite, pass rate and red-team results by version.
5. **Runtime guardrails.** Which guardrails apply in production, such as groundedness or PII redaction.
6. **Traffic path.** Whether model and tool calls pass through a gateway you control. This decides whether you can enforce or only observe.
7. **Cost per period.** Model spend attributed to this agent and version.
8. **Framework mappings.** Which controls map to OWASP, MITRE ATLAS, NIST AI RMF or ISO/IEC 42001 clauses.
9. **Last access review.** Date and reviewer. See [access reviews for AI agents](/blog/access-reviews-for-ai-agents).
10. **Dependencies.** Other agents it calls and agents that call it.
11. **Retirement criteria.** What would cause the agent to be switched off.
12. **Incident history.** Links to past incidents, quarantines and fixes.

## Which fields should discovery fill, and which need an owner?

The rule of thumb: if a field describes configuration, a scanner should fill it. If it describes intent, a person must.

- **Discovery fills:** platform, location, version, models, tools, MCP servers, credentials, traffic path, cost, last activity.
- **Identity provider fills:** owner status, team, leaver flags.
- **Owner fills:** purpose, user population, data classes, approval points, retirement criteria.
- **Reviewer fills:** risk tier confirmation, access review outcome.

Owner-supplied fields go stale. Keep them few, and ask for them once at registration with a short questionnaire. See [what an AI agent registration questionnaire should ask](/blog/ai-agent-registration-questionnaire).

## Should agents, models, tools and MCP servers be separate records?

Yes. Model each as its own record and link them. One MCP server is often used by many agents, and one model by hundreds. If the MCP server's permissions change, you want one record to update and a list of every agent affected. The same applies to keys: one leaked key should show every agent that used it. For the fields that belong on an MCP server record, see [MCP server inventory fields](/blog/mcp-server-inventory-fields), and for the full list of asset types see [which AI asset types to inventory](/blog/ai-asset-types-to-inventory).

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers. Auto-discovery across AWS, Azure and Google Cloud, plus connectors to Salesforce, SAP, Databricks, HubSpot and Claude Enterprise and OpenAI Enterprise workspaces, fills configuration fields without a survey. OIDC SSO and delegated on-behalf-of identity link owners and users to your identity provider. Eval results, LLM judge scores and cost per agent and per version attach to the record, and policy as code maps controls to OWASP, MITRE ATLAS, NIST and ISO 42001. [Book a demo](https://www.lyzr.ai/book-demo/).

For how the inventory fits into the wider governance layer, see [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## Frequently asked questions

**How many fields should an AI agent inventory have?**
Start with about 13 minimum fields and add an extended set later. More owner-supplied fields usually means a less accurate inventory.

**Can a CMDB hold an AI agent inventory?**
It can hold basic records, but agents change versions, tools and models often. Most teams need automated discovery feeding the record, with a link to the CMDB.

**What is the most commonly missing field?**
Owner. Agents built in pilots or by people who have since moved teams often have no accountable owner.

## Key facts

An AI agent inventory minimum field set: agent ID, name/description, platform and location, owner, owning team, business purpose, status, current version, models used, tools and MCP servers, credentials/identity, data access, risk tier. Extended fields: user population, on-behalf-of delegation, human approval points, eval history, runtime guardrails, traffic path, cost, framework mappings, last access review, dependencies, retirement criteria, incident history. Configuration fields should be filled by automated discovery; intent fields (purpose, data classes, approval points) by the owner. Agents, models, tools, MCP servers and keys should be separate linked records. Opencontroller by Lyzr keeps a governed registry of agents, versions, tools and MCP servers filled by auto-discovery across AWS, Azure, Google Cloud and connected platforms.

## Related

- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [Which AI asset types should an AI inventory cover?](https://opencontroller.lyzr.com/blog/ai-asset-types-to-inventory.md)
- [What should you record for every MCP server in your inventory?](https://opencontroller.lyzr.com/blog/mcp-server-inventory-fields.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
