What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
There is no single report that finds every AI agent in a workplace. Build the inventory by correlating platform and identity registries with SaaS-usage records, endpoint and network telemetry, and cloud workload inventories. For every finding, record who owns it, what it does, which identity and permissions it uses, what tools and data it can reach, and whether it is approved.
This layered approach separates a registered enterprise agent from a browser-based AI app, a locally installed agent, an unapproved “shadow” agent, and a custom workload running in the cloud. It also gives security teams enough context to approve, restrict, monitor, or retire each one.
As an Amazon Associate I earn from qualifying purchases.
What belongs in an AI-agent inventory?
Start by defining the population before collecting records. “AI agent” can mean an autonomous workflow that calls tools and takes actions, but employees may also use copilots, chat applications, scripts that call a model API, Model Context Protocol (MCP) servers, browser extensions, and internally built workloads. These categories have different risks and appear in different systems.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteKeep the categories separate in the registry. A chat application that only returns text should not be silently treated as an agent that can create tickets, change records, send mail, or move money. Record the broader AI application when it is relevant, then identify whether it has autonomous actions, scheduled execution, tool access, or delegated authority.
#1 Best Overall
Useful categories
- Registered platform agents: Agents created in an approved platform and represented by a platform or directory identity.
- Employee-facing AI applications: SaaS sites, desktop applications, browser extensions, and copilots used to generate or analyze content.
- Local agents: Software running on an employee workstation or server and making model or tool calls from that endpoint.
- Custom cloud workloads: Applications, models, orchestration components, and connectors deployed in a cloud account.
- Automation and integration code: Scripts, bots, and services that call models or tools even when they are not labeled “agents.”
- MCP servers and tool endpoints: Services that expose data or actions to an agent and therefore form part of the agent’s effective capability.
Why discovery requires several data sources
Each discovery surface sees a different population. A platform registry can describe an agent’s configuration but miss an employee’s local application. SaaS discovery can show that a user visited an AI service but usually cannot enumerate the tools and permissions of an agent runtime. Cloud inventory can reveal a deployed workload without showing which employee initiated its use.
Combine the sources and preserve the source and observation time for every record. This lets you distinguish “not found” from “not covered” and prevents an empty platform report from being mistaken for proof that no other agents exist.
| Discovery surface | What it can reveal | Typical blind spots or prerequisites |
|---|---|---|
| Agent and identity registries | Registered agents, identities, authentication, access controls, tools, knowledge sources, orchestration, endpoint context, and user context. | May omit unregistered, local, or externally hosted agents; visibility is tied to supported platforms and connectors. |
| AI-app and SaaS-usage discovery | AI sites and applications employees use, data interactions, app risk, and sanctioned or blocked status. | Shows application use rather than every runtime, agent configuration, or local tool. Device onboarding or a cloud-app connector may be required. |
| Endpoint and network telemetry | Processes, devices, users, destinations, and local agents making internet-bound connections; can help identify “shadow” agents. | Coverage depends on endpoint clients, forwarding policies, and—where required—TLS inspection. Traffic that never traverses the monitored path remains outside the view. |
| Cloud workload inventories | AI applications, models, components, dependencies, and AI bills of materials in cloud environments. | Does not by itself establish which employee uses a workload or whether a separate SaaS or endpoint agent is involved. |
Step 1: Define scope, ownership, and evidence rules
Write a short scope statement that answers four questions: which environments are included, which AI categories are included, which actions require registration, and how long observations remain useful. Include corporate endpoints, servers, cloud accounts, sanctioned SaaS tenants, and relevant development environments when they are in scope.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Assign an owner for the inventory itself and a responsible team for each record. Establish a minimum evidence standard, such as a platform record, an identity object, an endpoint process observation, a cloud resource, or a verified owner statement. Keep the evidence source and last-observed timestamp so an analyst can explain why a record exists.
Separate status from confidence
- Lifecycle status: Proposed, active, suspended, retired, or deleted.
- Governance status: Approved, conditionally approved, unapproved, or exception.
- Discovery confidence: Confirmed configuration, observed activity, inferred association, or owner-reported only.
- Coverage status: In scope and monitored, in scope but not yet covered, or explicitly out of scope.
These fields prevent an “unapproved” finding from being confused with an unverified lead, and prevent “not observed” from being read as “does not exist.”
Step 2: Start with registered agents and identities
In a Microsoft environment, begin with the Defender AI Agents page and the AgentsInfo Advanced Hunting table. Microsoft documents inventory entries for agents built with Copilot Studio, Microsoft Foundry, Microsoft 365, supported non-Microsoft platforms, and local agents discovered on endpoints. Depending on the source, records can include authentication, access control, tools, knowledge sources, orchestration, endpoint context, and user context.
Use Microsoft Entra Agent ID visibility as a separate identity-oriented view. It is designed to show identities created in Copilot Studio and Foundry. That makes it valuable for answering “which registered agents have identities?” but it is not a complete list of every AI application, local process, or externally hosted workload.
Free tools Windows power users keep installed
One-click scans. No signup required.
Make the current-state query historical-aware
The AgentsInfo table can contain multiple snapshots for the same AgentId. To build a current inventory, select the record with the newest timestamp for each identifier, then exclude records whose lifecycle state is deleted. Retain older snapshots in an audit store rather than overwriting them; version changes, permission changes, and deletion events are important evidence.
Rank #2
Local-agent entries have their own prerequisites. Microsoft documents Defender for Endpoint and Microsoft Defender Antivirus in active mode for local discovery, with local agents onboarded separately. Verify that the relevant endpoints are onboarded and that the local discovery feature is enabled before interpreting an empty result.
Step 3: Find employee use of SaaS AI applications
Use application-activity sources to identify AI services employees actually access. Microsoft Purview Data Security Posture Management for AI (DSPM for AI) provides visibility into AI activity and data interactions, including third-party SaaS applications such as ChatGPT Enterprise and Google Gemini. Monitoring third-party AI sites requires devices onboarded to Purview.
Defender for Cloud Apps can discover SaaS generative-AI applications, expose app-risk information, and support administrative sanction or blocking decisions. Treat these findings as application and usage evidence. A visit to an AI site does not prove that an autonomous agent ran, what tools it used, or which API calls it made.
Turn an app sighting into an inventory record
- Capture the application name, account or tenant, user, device, first-seen time, last-seen time, and source of the observation.
- Determine whether the service supports agents, custom instructions, connectors, API keys, scheduled jobs, or tool execution.
- Ask the user or application owner to identify the business purpose, data types handled, and any connected systems.
- Link the application record to a separate agent record when autonomous behavior or delegated access is confirmed.
- Apply the organization’s sanction, restriction, or exception workflow and record the decision.
Step 4: Detect local and shadow agents with endpoint and network telemetry
Microsoft Entra Global Secure Access AI agent discovery is documented as a preview dated June 2, 2026. It inspects internet-bound traffic and can attribute activity to the initiating user, device, and process. The service classifies agents as managed or shadow according to whether they are registered with Entra Agent ID.
For local-agent discovery, Microsoft lists three prerequisites: the Global Secure Access client, an internet access forwarding policy, and TLS inspection. Without those controls, the preview cannot provide the same local visibility. Confirm the feature’s current preview availability and tenant requirements before making it a dependency of an audit program.
The preview includes an inventory view and summary counts, including agents newly detected during the prior seven days. That seven-day figure is a reporting window, not a measurement of the percentage of agents detected. Use process, user, and device attribution to investigate an observation, then connect it to a registered identity or owner where possible.
Investigate a shadow-agent finding
- Confirm the process path, device, initiating user, destination, and observation time.
- Check whether the destination is an approved service and whether the process belongs to an approved application.
- Ask the user or team to identify the agent’s purpose, tools, data sources, and credentials.
- Compare its permissions with the minimum needed for the stated task.
- Register, restrict, isolate, or retire the agent according to policy, and preserve the original telemetry.
Step 5: Inventory deployed AI workloads in the cloud
Custom-built agents often appear as ordinary applications, functions, containers, model endpoints, queues, or connectors. Microsoft documents Defender for Cloud CSPM discovery for AI applications and models, including AI bill-of-materials information and Cloud Security Explorer queries. Use this source to enumerate deployed components and dependencies, then connect each workload to an owner, identity, environment, and business process.
Recommended Free Tools
Cloud discovery complements employee-usage and identity discovery. A workload may run on a schedule or service identity without a current interactive user, while an employee may invoke it through a separate portal. Record both the runtime identity and the human or team accountable for the workload.
Cloud checks that add context
- Identify model endpoints, orchestration services, storage locations, queues, functions, containers, and external connectors.
- Map service identities and role assignments to the tools and data the workload can reach.
- Record development, test, and production environments separately.
- Capture model and dependency versions, deployment timestamps, and change history.
- Link cloud resources to an approved application or agent record rather than creating duplicate names for every component.
Step 6: Design a registry that supports risk decisions
A name list is not an inventory. The registry must describe what an agent can do, who is accountable, and how its state changed. Microsoft guidance emphasizes ownership, purpose, platform, access scope, distinct agent identities, and observation of activity and tools. AWS Prescriptive Guidance recommends recording capabilities, permissions, purpose, ownership, version history, dependencies, performance and usage statistics, and approval and governance classifications.
| Field group | Record these values | Why it matters |
|---|---|---|
| Identity | Unique agent name and identifier; platform identity; service or user identity; authentication method; tenant and environment. | Prevents name collisions and shows whether actions run as a person or a non-human identity. |
| Accountability | Business owner, technical owner, responsible team, support contact, and escalation path. | Provides a person or team who can validate purpose and remediate the agent. |
| Purpose and process | Business process, intended users, autonomous actions, human approval points, schedule, and operating environment. | Allows reviewers to judge whether capability matches the stated need. |
| Capabilities and access | Permissions, data classifications, access scope, tools, APIs, MCP servers, knowledge sources, connectors, and write actions. | Defines the agent’s effective blast radius. |
| Technology | Platform, model or application components, deployment location, versions, dependencies, and endpoint context. | Supports change management, incident response, and reproducibility. |
| Lifecycle and evidence | Status, first seen, last observed, version history, source systems, activity links, and deletion records. | Distinguishes current state from historical snapshots and supports audits. |
| Governance and operations | Approval classification, exceptions, usage and performance statistics, cost metadata where available, review date, and retirement date. | Connects discovery to policy, capacity, and ongoing ownership. |
Step 7: Attribute every action to an identity
Do not let an agent execute under an anonymous shared credential when a distinct identity is possible. Microsoft and AWS guidance both support distinct agent identities, identity and authentication context, and audit trails. Record whether an action was initiated by a human, delegated to an agent, or performed by a service identity on a schedule.
Connect agent logs with directory, endpoint, API-gateway, application, and cloud audit records. The useful investigation chain is: human or triggering event, agent identity, device or workload, tool or API, target resource, result, and timestamp. If one link is unavailable, mark the gap explicitly rather than assigning the action to the nearest user.
Minimum audit questions
- Which identity authenticated to the agent and which identity authenticated to each tool?
- What user, schedule, event, or workflow started the run?
- Which model, prompt or instruction set, connector, API, and data source were used?
- What approval or human confirmation occurred before a consequential action?
- What records show the result, error, retry, or rollback?
Step 8: Reconcile findings with the approved registry
Compare every discovered record with the central registry. Investigate agents with no owner, no distinct identity, excessive permissions, unknown tools, or no approval status. An owner-backed decision should end each investigation: approve as-is, approve with conditions, restrict or block, migrate to a managed platform, or retire.
Reconciliation is not a one-time cleanup. Agent tools, model versions, connectors, and permissions can change without a new name. Schedule recurring comparisons and trigger reviews on material changes such as a new write-capable tool, a broader data scope, a new service identity, or a move from test to production.
Practical review cadence
- Continuous or daily: Ingest identity, endpoint, network, SaaS, and cloud observations.
- Weekly: Review new, shadow, unowned, and recently changed agents; the Global Secure Access preview’s seven-day “newly detected” view can help prioritize this queue.
- Monthly: Revalidate owners, permissions, tools, exceptions, and last-observed times.
- At release or change: Require review for new capabilities, integrations, model components, or production deployment.
- At retirement: Disable identities, revoke tokens, remove connectors, archive evidence, and mark the lifecycle record deleted or retired.
How to choose and combine discovery sources
Evaluate a source against the population it covers, the attribution it provides, configuration detail, prerequisites, governance actions, and freshness. No documented method covers every kind of workplace agent, so select a combination that matches your environment.
| Comparison axis | Questions to ask |
|---|---|
| Population | Does it identify registered platform agents, SaaS AI use, endpoint agents, deployed cloud workloads, or only one of these? |
| Attribution | Can a finding be tied to an owner, identity, user, device, process, or workload? |
| Configuration detail | Does it expose permissions, tools, knowledge sources, integrations, and lifecycle state? |
| Coverage prerequisites | Does it require endpoint onboarding, a connector, network forwarding, TLS inspection, or a particular cloud platform? |
| Governance action | Can administrators register, classify, sanction, block, remediate, or retire a finding? |
| Freshness and auditability | Can it separate current from historical snapshots and correlate activity with audit data? |
Common failure modes and fixes
“Our agent registry is empty, so we have no agents.”
Why it fails: A registry normally covers registered identities and supported platforms, not every local process, SaaS application, or cloud workload.
Fix: Add endpoint and network telemetry, SaaS-usage discovery, and cloud inventory, then label each source’s coverage.
“A SaaS access report tells us which agents ran.”
Why it fails: Application access proves use of a site or service, not autonomous execution, tool use, permissions, or API calls.
Fix: Link the app observation to agent configuration, identity, process, and audit records before classifying it as an agent.
“A name and owner are enough.”
Why it fails: Risk depends on capabilities, data scope, tools, identities, dependencies, and change history.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesFix: Require the full registry fields, especially permissions, connected tools, authentication method, and last-observed state.
“The network sensor covers every local agent.”
Why it fails: Local discovery depends on the documented client, forwarding policy, and TLS-inspection prerequisites, and traffic outside the monitored path is not visible.
Fix: Verify prerequisites per endpoint group and report uncovered populations explicitly.
“We can quote a detection percentage.”
Why it fails: The official guidance reviewed here does not provide a neutral, comparable detection-rate benchmark for all workplace agents.
Fix: Report coverage by source, population, prerequisite, and observation window instead of claiming universal detection.
What a defensible inventory looks like
A defensible inventory can answer, for any record, what the agent is, who owns it, why it exists, where it runs, which identity it uses, what it can access, which tools and APIs it calls, whether it is approved, and when those facts were last observed. It can also show what the discovery method could not see.
Use the inventory as a control plane rather than a spreadsheet snapshot. Feed it from platform, identity, endpoint, network, SaaS, and cloud sources; preserve historical observations; and route exceptions to an owner. This is the practical path to finding both sanctioned agents and the unregistered or “shadow” agents that a single platform view will miss.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
