---
title: What questions reveal whether an AI agent discovery tool really covers your estate?
url: https://opencontroller.lyzr.com/blog/questions-to-test-ai-agent-discovery-coverage
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What questions reveal whether an AI agent discovery tool really covers your estate?

**Don't judge discovery by a demo of what it found. Plant known agents, models, API keys and MCP servers across your clouds, SaaS platforms and laptops, then measure how many the tool finds, how fast, and with what detail. Ask vendors about read permissions, scan frequency, and what each connector can see versus only list.**

Every discovery tool finds something in a demo. The question is what it misses in your estate. A planted-asset test answers that in about two weeks.

## What should you ask before the test?

Ask these in writing. Vague answers are a signal.

1. **Which surfaces does it cover?** Name each: AWS, Azure, Google Cloud, each agent platform you use, enterprise AI workspaces, developer laptops and browsers.
2. **What permissions does each connector need?** Read-only roles should be enough. Compare with our guide to [discovery permissions on AWS, Azure and Google Cloud](/blog/ai-agent-discovery-permissions-aws-azure-google-cloud).
3. **How often does it scan, and what triggers a rescan?** See [how often AI agent discovery should run](/blog/how-often-should-ai-agent-discovery-run).
4. **Which asset types does it return?** Agents, models, deployments, API keys, MCP servers, skills, tools, knowledge sources. Our list of [AI asset types to inventory](/blog/ai-asset-types-to-inventory) is a good baseline.
5. **What does "found" include?** Name only, or owner, model, tools, identity, last activity and cost?
6. **How does it deduplicate?** The same agent can appear in a cloud account, a gateway log and a SaaS admin API.
7. **What can it find without admin API keys?** Some platforms need admin consent; ask for fallbacks. See [discovering AI use without admin API keys](/blog/discover-ai-usage-without-admin-api-keys).

## How do you run a planted-asset coverage test?

Plant 20 to 40 known assets before the trial starts. Record each in a sheet with its location and creation time.

| Surface | Example planted assets | What a good result shows |
|---|---|---|
| AWS | A Bedrock agent, a model invoked from a Lambda, an IAM access key used for model calls | Agent, model, calling identity, Region |
| Azure | A Foundry agent, a model deployment, a key-based OpenAI-compatible endpoint | Agent, deployment, key vs Entra ID auth |
| Google Cloud | A Vertex AI agent, a model endpoint in a second project | Agent, project, model, service account |
| Agent platforms | An agent in each SaaS platform you use, one created by a non-admin | Agent, creator, connected data sources |
| AI workspaces | A custom GPT or project, a skill, a connector | Item, owner, sharing scope |
| Developer devices | A coding agent and a local MCP server on a test laptop | Device, user, server command or URL |
| Keys and gateways | A model provider key in a CI secret, a self-hosted LLM proxy | Key owner, last use, unknown gateway flagged |

Include two hard cases: an agent created in a sandbox account nobody registered, and an agent deleted halfway through the test. The first tests reach. The second tests whether inventory goes stale.

## How do you score the result?

Measure four numbers for each surface:

1. **Recall:** planted assets found ÷ planted assets. Aim for every planted asset on surfaces the vendor says it covers.
2. **Time to find:** hours from creation to appearance in inventory.
3. **Detail completeness:** share of found assets with owner, model, tools and identity populated.
4. **False or duplicate entries:** items that are not real assets or appear twice.

Then score what the tool found that you did not plant. Real unknown agents are the best evidence of value, but check each one by hand before counting it.

Finally, test the step after discovery. Pick one found asset and ask the tool to show what you can do: assign an owner, register it, apply a policy, or block it. Registration is not runtime control; ask which actions work on each surface. Our [shadow agent playbook](/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it) covers that step, and the [platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist) covers the rest of the evaluation.

## Which red flags should end the evaluation?

- Coverage claims that drop when you name a specific platform.
- Write or admin permissions required for read-only discovery.
- No answer on deduplication.
- Found assets with no owner or identity field.
- Discovery that only runs on a manual trigger.

Inventory is the first control in most frameworks. CIS Controls v8 starts with inventories of enterprise assets and software, and NIST CSF 2.0 places asset management under its Identify function. A tool that misses assets weakens every control after it.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud. Connectors cover third-party agent platforms (Salesforce, SAP, Databricks, HubSpot) and Claude Enterprise and OpenAI Enterprise workspaces, and it discovers and manages skills in Anthropic Enterprise. Opencontroller Edge extends discovery to developer devices, coding agents and local MCP servers, installed by IT via MDM. Found assets go into a governed registry. Enforcement depth depends on the connector and traffic path, from registry to observe to enforce. Read [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents), see the [capability reference](/capability-reference), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How many assets should we plant in a discovery test?**
Twenty to forty across every surface you use is enough to see gaps without slowing the trial.

**What recall should a discovery tool reach?**
It should find every planted asset on the surfaces the vendor says it covers. Misses there are product gaps, not edge cases.

**Should discovery need admin permissions?**
No. Discovery should work with read-only roles. Ask why if a connector needs write access.

## Key facts

To test AI agent discovery coverage, plant 20-40 known assets (agents, model deployments, API keys, MCP servers, skills, coding agents) across AWS, Azure, Google Cloud, SaaS agent platforms, AI workspaces and developer laptops, then measure recall, time to find, detail completeness and duplicates per surface. Discovery should need only read-only permissions. Opencontroller (by Lyzr) auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, connects to Salesforce, SAP, Databricks, HubSpot, Claude Enterprise and OpenAI Enterprise, and extends to devices with Opencontroller Edge via MDM. Pricing starts from $250K per deployment per year.

## Related

- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [Which AI asset types should an AI inventory cover?](https://opencontroller.lyzr.com/blog/ai-asset-types-to-inventory.md)
- [How should you run a pilot of an AI agent governance platform?](https://opencontroller.lyzr.com/blog/how-to-run-a-pilot-of-an-ai-agent-governance-platform.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
