Lyzr Research · The AI Estate Inventory Playbook · November 2026
The AI Estate
Inventory Playbook.
How to find every agent your enterprise runs, reconcile what people say against what systems see, and keep the register true. Twelve hiding places, eight lenses and a ten day sprint.
To CISOs, CIOs, heads of AI, internal audit and the people given the job of counting
A list of agents is not an inventory. A ledger is.
Most inventories are lists of claims.
A spreadsheet that someone filled in from memory says what people believe. It carries no evidence, no date and no owner who has signed. It decays the week after it is made.
A ledger holds each claim against evidence.
Every line says how we know it exists, who answers for it, how much it can do and when we last checked. Auditors have worked this way for a century, and the same tests apply to agents.
You can build the first one in ten working days.
This playbook gives the places to look, the lenses to look through, the fields to record, the way to find owners and a day by day sprint. It ends with how to keep the register true.
This is the method
Every figure is cited. Most telemetry comes from security vendors and is labeled as such. Part 10 is the one place we write about Lyzr, and it is marked.
Lyzr Research, November 2026
The short version.
- The ledgerTreat the register as a ledger of claims. Grade each line as said, seen, shown or shut.
- The placesAgents hide in twelve places. Most teams watch about four.
- The lensesNo single scan finds them all. Use eight lenses and merge them into one record.
- The overlapReconcile said against seen. The gaps are Known, Stray, Ghost and Orphan.
- The registerStart with six fields. Six complete beats eighteen half done.
- The ownersWalk a ladder to find owners, offer amnesty, then tier by blast radius.
- The sprintTen working days from scope to readout, with a date to count again.
- The pulseAdd triggers, a front door gate and five vital signs, or the register decays.
What this covers
Ten parts, from the ledger to the first census.
What a register is
The ledger
Why a list fails, and the evidence grades that make a ledger.
Where agents hide
Twelve places
A field guide to the places a census must visit.
How to look
Eight lenses
What each discovery method sees, and what it misses.
Said against seen
Reconcile
Four classes of finding and a way to size the unseen.
What to record
Register design
Six fields first, IDs, states and design failures.
Who answers
Owners and triage
The owner ladder, amnesty and blast radius tiers.
Do it
The ten day sprint
Day by day, with roles and a worked example.
Stay true
Keep it alive
Triggers, the front door gate and vital signs.
What is asked
Frameworks and rules
NIST, ISO 42001, the EU AI Act and OWASP.
A vendor view
Where Lyzr fits
Labeled. Where a platform can help.
The argument in one picture
A register is a ledger of claims, not a list of names.
A list says an agent exists. A ledger says how we know, who answers for it, and how recently someone checked. The more evidence on a line, the harder it is to fake and the longer it stays true.
How to use this playbook.
The ledger.
What a register is, and why a list is not enough.
This part sets out the one idea the rest depends on. A register is a set of claims, and a claim without evidence is only an opinion. It introduces four evidence grades and six audit tests.
Four grades of evidence
Said, seen, shown, shut.
Said means a person told you. Seen means a system observed it. Shown means a named owner confirmed it. Shut means a control enforces it. Report each count on its own, because a register that is all said is the weakest kind.
What each grade looks like
Four grades, four kinds of proof.
Said
A survey answer, an email, a line in a spreadsheet. Cheap to get and quick to go stale.
Seen
A scan hit, a log line, an entry in an admin console. Objective, but silent about who owns it.
Shown
An owner has confirmed it, in writing, with a date. Slower, and it is what makes a record accountable.
Shut
A control makes the record a condition of working, such as no credentials without registration.
Our synthesis, offered as a starting point. The graphics are illustrative.
IBM Think 2026, via Alatirok
Fewer than one in five hold a complete inventory.
Share of organizations reporting a complete AI inventory
Most organizations have something. Few have everything. The gap between a partial list and a full ledger is where unowned agents live.
Survey figure, as cited by a third party.
IBM Think 2026, as cited by Alatirok. See also the Cloud Security Alliance shadow AI note.
Six tests for a register
Auditors already know how to test a ledger.
Existence, completeness, rights, classification, valuation and cut-off are the assertions auditors apply to any record. They map one to one onto an agent register, so the same discipline works here.
The tests, in plain words · 1 of 2
The six tests, applied to agents.
Existence
Pick ten lines. Can you show each agent running today, with a log or a screenshot.
Completeness
Pick ten agents from a scan. How many are on the register. The miss rate is the answer.
Rights
Does a named person, not a team alias, answer for each line, and have they accepted it.
ISA 315 and ACCA audit assertions, adapted by Lyzr Research.
The tests, in plain words · 2 of 2
The six tests, applied to agents.
Classification
Are tiers set by what the agent can do, not by its name or its vendor.
Valuation
Is the blast radius written down for every agent above the lowest tier.
Cut-off
Is each record dated, and was it confirmed inside the review window.
ISA 315 and ACCA audit assertions, adapted by Lyzr Research.
The principle
“Count the evidence, not only the agents.”
THE LEDGER RULE
Report how many lines are said, seen, shown and shut. A register of 400 agents at 90 percent said is a smaller fact than a register of 120 agents at 80 percent shown.
Why lists fail
Why a list fails inside a quarter.
No evidence
Nothing on the line says how we know. The next person cannot tell a fact from a guess.
No owner who signed
A team name is not a person. When something goes wrong nobody picks up.
No date
Without a last confirmed date, a stale record and a live one look identical.
No trigger
Nothing forces an update when an agent is built, changed or retired.
No link to action
A register that cannot change an access decision is a document, not a control.
Our synthesis, informed by the Anomity and enz.ai inventory guides.
Twelve places.
Where agents hide in a real enterprise.
Agents are not only built in the AI platform. They are switched on in SaaS settings, wired up in automation tools, installed on laptops and granted access through personal accounts. This part is the field guide.
The twelve doors
Four doors are watched. Eight often are not.
Most teams monitor the cloud AI platform, the main SaaS suites and the low-code builder they bought. The other doors lead to agents that nobody reviewed, and they are where a census earns its keep.
Places one to six
Places one to six.
Cloud AI platforms
Agents built in a cloud console by one team, in an account central IT may not review.
Embedded SaaS agents
Assistants and agents switched on inside tools the business already pays for.
Low-code builders
Business users assemble agents with drag and drop, often with broad connectors.
Workflow automation
Automation flows with a model call in the middle and credentials in the settings.
Coding assistants
Assistants that read repos, run commands and call tools from a developer machine.
Local MCP servers
Tool servers running on laptops that give assistants access to files and services.
Our synthesis, offered as a starting point.
Places seven to twelve
Places seven to twelve.
Repos and notebooks
Scripts and notebooks that call a model API with a key in code or in the environment.
CI/CD pipelines
Pipelines that call models to review, test, triage or deploy.
Browser and desktop agents
Extensions and desktop tools that act on a user’s behalf in other applications.
Personal OAuth grants
A user links an outside agent to work mail, drive or calendar with one click.
Vendor and contract AI
AI that arrived in a vendor product update or a services contract, never reviewed as AI.
Data platform agents
Agents inside warehouses and lakehouses that query and write data directly.
Our synthesis, offered as a starting point.
Clutch Security, December 2025, vendor data
In a modeled 10,000 person organization, about 1,500 people run MCP servers.
Share of employees running MCP servers in Clutch’s model for a 10,000 person organization
Clutch analyzed more than 15,000 MCP deployments and modeled the result at about 3,056 deployments in one organization. 38% of those were unofficial, and 86% ran locally. Treat the exact figures as a vendor model, and the direction as the point.
Vendor data. Modeled, not measured in one customer.
Clutch Security, December 4, 2025. Vendor research, sample size beyond 15,000 deployments.
What the MCP data says
Most MCP servers are local, unofficial and on endpoints.
The shape matters more than the decimal. Local servers on laptops are invisible to cloud scans, which is why endpoint and code lenses are needed. Clutch also found 3% of published servers held valid hardcoded credentials.
Pluto Security, August 2026, vendor data
Across four customer environments the count of distinct MCP servers grew by 72% to 98% in three months.
Median distinct MCP servers per organization in Pluto’s sample
From May to August 2026, Pluto saw distinct MCP servers on endpoints grow 72% to 98%. The median organization had 288 distinct servers and the median endpoint had 3. Sample sizes are mostly not stated, so read this as direction.
Vendor telemetry, four environments.
Pluto Security, August 2026. Vendor telemetry.
Cyberhaven Labs, vendor data
Native AI apps on endpoints grew by about five times in a year.
Growth in AI native apps on endpoints, February 2025 to February 2026
Coding assistants grew 357% over the same year, across hundreds of thousands of employees. Endpoints are where a lot of the estate now lives, and they are the lens most programs lack.
Vendor data.
Cyberhaven Labs, 2026. Vendor telemetry.
Why each place is missed
Why each place gets missed.
Different owners
Cloud teams, SaaS admins, developers and procurement each own one door and none sees the others.
No shared vocabulary
Platform vendors call them assistants, copilots, flows, skills or servers. A scan for agents finds only one word.
Silent defaults
Some features arrive switched on in an update, so no project ever existed to be listed.
Personal accounts
Consent grants and personal plans sit outside the corporate tenant.
Nothing to approve
If setup takes a minute and needs no ticket, there is no ticket to find later.
Our synthesis.
Eight lenses.
How to look, and what each way of looking misses.
Discovery is a set of lenses, not a single scan. Each lens sees some doors well and others not at all. This part describes eight, with the signals to look for and the blind spot of each.
Eight lenses, one record
Eight ways of looking feed one register.
Cloud accounts, SaaS consoles, code and CI, endpoints, identity, network egress, spend and people. Each produces a list. A census merges them into one record and notes which lens saw which agent.
Lens one and two
Lenses one and two, the platforms.
Cloud accounts
List agent, model and registry resources in every account and project. Include the accounts nobody on the central team created. Blind spot, everything off the cloud.
SaaS consoles
Export agents, assistants and automations from each admin console and builder. Ask each vendor what is on by default. Blind spot, tools without an admin API.
Our synthesis, informed by the Anomity discovery domains guide.
Lens three and four
Lenses three and four, the builders and the laptops.
Code and CI
Search repos for model SDK imports, API key patterns and workflow files that call models. Search for MCP config files such as .cursor/mcp.json, claude_desktop_config.json and .vscode/mcp.json. Blind spot, code outside your org.
Endpoints and browsers
Inventory installed AI apps, extensions and local servers through endpoint management. Look for listening local ports. Blind spot, unmanaged and personal devices.
Our synthesis. Config file names are common examples and vary by tool.
Lens five and six
Lenses five and six, the traffic and the grants.
Identity and OAuth
List app registrations, consent grants, service accounts and token scopes. Flag any grant to an AI vendor. Blind spot, agents that use shared secrets rather than identities.
Network egress
Match proxy and DNS logs to model provider domains and gateways. It reveals use even when nothing is installed. Blind spot, encrypted tunnels and personal networks.
Our synthesis.
Lens seven and eight
Lenses seven and eight, the money and the people.
Spend and procurement
Search card statements and purchase orders for AI vendors and usage billing. Spend finds what nothing else does. Blind spot, free tiers.
People and amnesty
Ask. A short survey with an amnesty promise finds agents that were never recorded anywhere. Blind spot, whatever people forget or hide.
Our synthesis.
What each lens sees and misses.
How many lenses
Run at least four lenses, and always one of each kind.
- One platform lens, from a cloud or SaaS console.
- One builder lens, from code and CI.
- One traffic or grant lens, from identity or egress.
- One human lens, from people and amnesty.
Our recommendation. Independent lenses also allow the estimate in Part 04.
Scan hygiene
Five rules for scanning.
Read only first
Start with read-only access. A census that changes things loses trust on day two.
Write the definition first
Agree what counts as an agent before the first query, then freeze it for the sprint.
Log the query
Record each query and date so the evidence can be reproduced by an auditor.
Name the blind spot
State what each lens could not see in the readout.
Do not rate yet
Collect before you judge. Tiering is a later step.
Our synthesis.
The definition
One definition of an agent, written before you scan.
Autonomy
It can take actions or call tools without a human approving each step.
Context
It holds instructions, memory or retrieved data beyond a single prompt.
Reach
It is connected to at least one system, file or service that can change or leak something.
Edge cases
A chat window with no tools is out of scope. A workflow with a model call and write access is in.
Our synthesis. Adjust to your policy, and keep it stable during a count.
Reconcile.
What people said against what systems saw.
A census produces two lists. One is what people say they run. The other is what systems observed. The gaps between them are the findings, and they come in four kinds.
Four classes of finding
Known, Stray, Ghost and Orphan.
Known is said and seen. Stray is seen but nobody said it. Ghost is said but never seen. Orphan is seen and recorded, with no live owner. Each class needs a different next step.
What to do with each class
A different move for each class.
Known
Confirm the owner, grade the evidence, tier it. Most of the register should end up here.
Stray
Find the owner using the ladder in Part 06. Then register it, or retire it. Strays are the main risk.
Ghost
Check whether it exists. It may be retired, renamed or running in a place no lens covers. Close or locate it.
Orphan
The owner left or changed role. Assign a new named owner within a set number of days or disable access.
Our synthesis.
Identity resolution
Joining the lists is the hard part.
Pick join keys
Use a stable ID where one exists, then the account, the workspace, the credential ID, the owner email and the endpoint name.
Expect many names
The same agent will be called different things by a cloud console, a repo and a person. Match on keys, not on names.
Keep the link
Store which lens and which record produced each match, so the result can be audited.
Mark uncertainty
Mark a match as probable or confirmed. Do not force a clean join.
Review the unmatched
The unmatched are where the findings are. Review them by hand.
Our synthesis.
Sizing the unseen
A way to estimate what you have not found.
If two lenses look independently, the overlap tells you how much each one misses. Method A found 40, method B found 30 and 12 appear in both, so about 100 exist and about 42 remain unseen. It is an estimate, and it is honest about the gap.
When the estimate holds
Use the estimate with care.
Independence
The lenses must look in different ways. Two scans of the same cloud are not independent.
Same definition
Both lenses must count the same thing, at the same time.
Heterogeneity
Easy to find agents are over-represented in both lists. That means the true total is likely higher.
Report a range
Say about 100, probably more. Do not print a single number as if it were a count.
Capture and recapture is a standard population estimating method. Adapted by Lyzr Research. An illustrative example.
Reco 2025, via the Cloud Security Alliance, vendor data
Shadow AI tools can run for more than a year before anyone detects them.
Median days to detect shadow AI tools in the cited telemetry
A single census pass is a snapshot. The longer an agent runs unseen, the more data and access it accumulates. This is why reconciliation has to repeat.
Vendor telemetry, as cited by a third party.
Reco 2025, via Cloud Security Alliance AI Safety Initiative, Shadow AI Apps, May 30, 2026.
A reconciliation worksheet, one row per finding.
A good reconciliation
What a good reconciliation looks like.
- Every finding has a class and a due date.
- The count of unmatched on both sides is reported, not hidden.
- Each join is explained and can be replayed.
- The estimate of the unseen is stated with its assumptions.
Our synthesis, offered as a starting point.
Register design.
What to record, and how to name it.
Registers fail from over-design more than from under-design. This part sets out the six fields that make a record useful on day one, the next fields to add, how to assign IDs and the states an agent passes through.
Six fields first
A register with six fields beats one half done in eighteen.
Name, owner, place, reach, autonomy and evidence. If you cannot fill these for an agent, you cannot make a decision about it. Add more fields only when somebody uses them.
The six fields in detail.
Fields to add second · 1 of 2
The next ten fields, in the order people ask for them.
Purpose and users
What it is for and who it serves, in a sentence.
Model and version
Which models it calls, and through whom.
Data classes
The sensitivity of what it handles.
Tools and MCP servers
Every tool it can call.
Human approval points
Steps that need a person.
Our synthesis.
Fields to add second · 2 of 2
The next ten fields, in the order people ask for them.
Vendor and contract
Who supplies it and under what terms.
Tier
Output of the triage in Part 06.
Last review
The date of the last confirmation.
Retire by
A date or a condition.
Links
Tickets, repos, runbooks, dashboards.
Our synthesis.
IDs and states
IDs never change. States do.
Issue one ID per agent
Do not reuse an ID, ever. A retired agent keeps its ID so history stays readable.
Keep external IDs as aliases
Cloud resource IDs and vendor IDs are aliases, since they change with migrations.
Use five states
Proposed, Active, Paused, Retired and Unknown. Unknown is for findings not yet classed.
Allow only listed moves
Proposed to Active needs an owner and a tier. Active to Retired needs a revocation record.
Our synthesis.
Test the register
Run the six tests on your own register before you publish it.
Existence, completeness, rights, classification, valuation and cut-off apply to the register itself. If it fails any of them, say so in the readout. A flawed register that admits it is more useful than a clean one that does not.
Design failures · 1 of 2
Six ways registers fail by design.
Defining AI too narrowly
Counting only models you built misses embedded and vendor agents.
One function owns it
A register owned only by security lacks the build knowledge. Joint ownership with engineering works better.
Waiting for perfection
Publish version one in the sprint. A rough register in use beats a perfect one in draft.
Failure modes drawn from the enz.ai inventory guide, a vendor source. The decay claim is that guide’s, not measured here.
Design failures · 2 of 2
Six ways registers fail by design.
Ignoring vendor AI
Contract AI is part of the estate even if nobody on your team built it.
Disconnected from action
If nothing in access, procurement or change management reads the register, it will decay.
No maintenance plan
One guide suggests an inventory may be materially incomplete within three to six months.
Failure modes drawn from the enz.ai inventory guide, a vendor source. The decay claim is that guide’s, not measured here.
Owners and triage.
Put a name beside every agent, then decide how much care each needs.
An unowned agent cannot be reviewed, retired or defended. This part gives a ladder for finding owners, an amnesty window, and a four axis blast radius score that sets the tier.
Entro Labs, as cited by the Cloud Security Alliance, vendor data
Non-human identities may outnumber people by 144 to 1.
Ratio of non-human to human identities reported in one vendor dataset
Rubrik reports about 45 to 1 in another dataset. Whatever the exact ratio, most identities are not people, and ownership is not obvious.
Vendor data. Ratios vary widely.
Entro Labs and Rubrik, as cited by Cloud Security Alliance AI Safety Initiative, Shadow AI Apps, May 30, 2026.
The owner ladder
Walk down the ladder until someone answers.
Start with the creator of record, then the last editor, the cost center, the creator’s manager and the team lead. If none answer, the amnesty desk takes the agent and starts a clock. Never stop at a team alias.
Amnesty
Offer amnesty, and mean it.
The offer
For two weeks, anyone who registers an agent they built outside process will not be penalized for that.
The condition
After the window closes, an unregistered agent found by a scan is handled as a finding.
The message
Send it from a senior sponsor, not from security alone. State the benefit, which is support and safe credentials.
The channel
Use a one page form and a named inbox. If registering takes more than ten minutes, people will not do it.
Our synthesis.
Cloud Security Alliance and Oasis, vendor commissioned
About half of organizations say agent ownership is unclear.
Share reporting no clear owner for AI agents or their identities
Ownership is not a data entry problem. It is a decision about who is accountable, and it needs a sponsor.
Survey, as cited by a third party.
CSA and Oasis, as cited by the Anomity inventory guide.
Blast radius
Score reach, power, autonomy and exposure.
Reach is how many systems and people it touches. Power is what it can change or spend. Autonomy is how often it acts alone. Exposure is what outsiders can reach or feed it. Score each from one to five.
The lethal trifecta
Three properties that should never meet unguarded.
Private data
The agent can read information that outsiders should not see.
Untrusted content
It reads text from email, web pages, tickets or documents that anyone can write.
External action
It can send, post, call or move something outside.
Why it matters
With all three, hidden text in untrusted content can make an agent send private data out. Remove one leg.
Simon Willison, The lethal trifecta for AI agents, June 16, 2025.
Tiers set the care, not the name.
Using the tiers
How tiers change what happens.
Credentials
Higher tiers get shorter lived, narrower credentials.
Approvals
Tier 3 and 4 agents need a human to approve irreversible steps.
Monitoring
Tier 4 agents are watched continuously, and anomalies page a person.
Retirement
Any agent unused for 90 days moves to Paused, and its credentials are revoked.
Our synthesis.
Triage in one page
What the triage page must show.
- Every agent, its tier and its owner.
- The ten highest scores and who is acting on each.
- Agents with no owner after amnesty, and the date they will be paused.
- The count of strays and orphans open, and the age of the oldest.
Our synthesis, offered as a starting point.
The ten day sprint.
From scope to executive readout, in two working weeks.
A first census does not need a program. It needs ten days, six roles and a plan. This part sets it out day by day, with the output of each day and the trap to avoid.
Ten days at a glance
Ten working days. Every bar produces a file.
Scope on day one, pull lists from clouds and SaaS on days two and three, scan code and endpoints, reconcile on day five, run amnesty and owner notices across the second week and brief the executive sponsor on day ten.
Days one to three, scope and collect.
Days four to six, widen and reconcile.
Days seven to ten, close and report.
Who does what · 1 of 2
Six roles, none of them full time.
Sponsor
A senior executive who sends the amnesty message and clears blockers.
Program lead
Runs the sprint, owns the plan and the readout.
Cloud and SaaS admins
Pull the lists from the platforms they run.
Our synthesis.
Who does what · 2 of 2
Six roles, none of them full time.
Engineering lead
Scans repos and pipelines and answers for what the scans find.
Identity and network leads
Supply grants, service accounts and egress matches.
Risk or audit partner
Runs the six tests and signs off the method.
Our synthesis.
A worked example, composite
A composite sprint, in numbers.
Illustrative composite, not a real company. Four lenses produced 212 distinct candidates. 96 were Known, 71 Stray, 28 Ghost and 17 Orphan. After amnesty, 58 strays gained owners and 13 were retired. Evidence moved from 20% shown to 52% shown.
The composite, by class.
The sprint traps
Five ways a sprint slips.
Debating the definition
Fix it on day one and revisit it only after the sprint.
Waiting for admin access
Request access before day one, not on day two.
Over-scoring
Scoring everything slows the week. Score strays first.
Skipping the readout
No readout means no decision, and no next count.
Treating it as one-off
Book the second census before the first one ends.
Our synthesis.
The executive page
The executive readout fits on one page.
The number
How many agents, in which classes, with the range for the unseen.
The grades
How much of the register is said, seen, shown and shut.
The risk
How many are Tier 3 or 4, and how many have no owner.
The ask
Two or three decisions needed from the sponsor.
The date
When the next count will be done.
Our synthesis.
Keep it alive.
A register is true on the day it is made. Then it starts to decay.
The first census is the easy part. This part covers what keeps the register true, which is update triggers, a front door gate, five vital signs and a plan for the registers that platforms now ship.
Why inventories rot
Without triggers an inventory drifts away from the estate.
New agents appear weekly, owners change role and tools update themselves. A register held up by goodwill decays. One held up by triggers and a gate stays close to the truth. The curves are illustrative, not measured.
Update triggers · 1 of 2
Six events that must update the register.
A new agent is built or bought
Registration is part of the build or purchase checklist.
An owner changes role or leaves
HR events trigger an ownership check.
A new tool or MCP server is added
A change in reach is a change in tier.
Our synthesis.
Update triggers · 2 of 2
Six events that must update the register.
A vendor ships an AI feature
Procurement flags renewals and release notes.
An incident or near miss
Re-score the tier and re-confirm the owner.
A scheduled recount
Run lenses again every quarter, and the full sprint yearly.
Our synthesis.
The front door gate
No registration, no credentials.
The most reliable control is one the builder needs. If an agent cannot get a credential, a key or a connector grant until it has a register entry, registration becomes the path of least resistance. This is how a record moves from shown to shut.
Building the gate
Where to place the gate.
Secrets and keys
Issue model and tool credentials only against a registered agent ID.
OAuth consent
Require admin approval for AI vendor consent grants, with a registration check.
Procurement
No purchase order for AI features without a register entry.
CI/CD
A pipeline that calls a model must reference a registered ID.
Exceptions
Allow a time limited exception with a named approver, and log it.
Our synthesis.
Five vital signs
Measure coverage, freshness, ownership, evidence and speed.
Coverage is the share of the estate with a record. Freshness is the share confirmed in the last 90 days. Ownership is the share with a live named owner. Evidence is the share at shown or shut. Speed is how soon a new agent is registered. Figures shown are illustrative.
What each vital sign tells you.
Microsoft Agent 365
Microsoft documents a registry that sees more than Entra does.
Two views
The Entra admin center lists only agents that hold an Entra Agent ID. Agent 365 is described as the unified registry and control plane.
A broader inventory
Agent 365 inventory includes non-Microsoft agents and agents without an Entra Agent ID.
Where to look
Admins view it under Agents, All agents, in the Microsoft 365 admin center, with at least the AI Reader role.
The catch
Coverage depends on what is connected. Treat it as one feed, not the estate.
Microsoft Learn, Agent registry convergence, updated 2026. Vendor documentation.
The Register, April 9, 2026
AWS added a registry in preview across five regions.
Regions where AWS Agent Registry launched in preview through Bedrock AgentCore
It catalogs agents, tools, MCP servers and agent skills, supports MCP and A2A, and records can be added manually or fetched from configured endpoints. Preview features change, so check the current status before relying on it.
Vendor announcement, preview.
The Register, April 9, 2026.
A register of registers
Platform registries are feeds. Your register is the whole.
Treat each platform registry as one input, alongside code scans, identity and spend. Pull from them on a schedule, keep the source on each record and resolve duplicates against your own IDs. No single vendor sees the rest.
Registry choices
Choosing where the system of record lives.
Use a platform registry
If most of your agents run on one platform, its registry is the natural first feed. Keep your own ID layer on top.
Use a governance platform
If you run many platforms, a neutral layer that ingests all the feeds saves manual joins.
Use a plain database
A governed spreadsheet or table is enough for the first six months if the fields are enforced.
Whatever you choose
Make it read by access, procurement and change management, or it will not stay current.
Our synthesis.
Operating model
Who runs the register after the sprint.
An owner for the register
One named lead with time and a mandate, not a shared inbox.
A monthly review
Thirty minutes on the vital signs, the open strays and the oldest orphan.
A quarterly recount
Repeat the four lenses and refresh the estimate of the unseen.
An annual sprint
Run the full ten days again, and compare with the last.
Our synthesis.
Failure modes · 1 of 2
Six ways a living register dies.
The champion leaves
Nobody else has the mandate.
The gate has a back door
People find a path around registration.
Triggers are manual
No system tells the register that something changed.
Our synthesis.
Failure modes · 2 of 2
Six ways a living register dies.
Metrics are vanity
Counting records instead of confirming them.
The register is private
If only security reads it, nobody else updates it.
Scope creep
Adding fields until updating takes an hour.
Our synthesis.
Frameworks and rules.
What the standards and the law ask you to know.
A register is not a compliance exercise, but it answers many compliance questions at once. This part maps the register to NIST, ISO 42001, the EU AI Act and the OWASP Top 10 for Agentic Applications.
NIST AI RMF, Govern 1.6.
ISO 42001 Annex B, resource documentation.
The EU AI Act dates a register helps with.
OWASP in one view
Ten agentic risks, each one needs a register field to address it.
OWASP published its Top 10 for Agentic Applications in December 2025, with input from more than a hundred researchers. Most of the ten are easier to manage when the register says what an agent can reach and who owns it.
OWASP risks one to five, and the field that helps.
OWASP risks six to ten, and the field that helps.
One register, many answers
What one register answers.
- Which agents does a regulator or customer ask about, and who owns each.
- Which controls apply to which tier, and the evidence they hold.
- Which risks from the OWASP list each agent is exposed to.
- Which agents moved this quarter, and why.
Our synthesis, offered as a starting point.
Where Lyzr fits.
A vendor view, clearly labeled.
Everything before this part is vendor neutral and you can run it with a spreadsheet. This part is the one place we write about Lyzr, and what we think a platform can add.
Where a platform helps and where it does not.
Lyzr Opencontroller, one data point
In one enterprise, Opencontroller surfaced 227 shadow assets.
Shadow assets surfaced in one enterprise by Lyzr Opencontroller
One data point, not an average. Your number will differ, and a census will tell you what it is. We include it only to show how large the gap between known and actual can be.
Single customer observation. Not a benchmark.
Lyzr, internal observation.
Free Lyzr tools that pair with this playbook.
About Lyzr
The full-stack agent infrastructure platform for enterprises and governments that want to own their intelligence.
Founded in New York in 2023. Agent Studio, Architect, Opencontroller, Open Cowork, Openfinetuner, Optimus and Madison cover building, governing, running and owning agents. Deploys in your AWS, Azure or GCP account, a dedicated tenant, SaaS, on bare metal, fully air-gapped, or on the Optimus appliance.
Backed by Rocketship.vc, BGV, Accenture Ventures, GFT, the Partnership Fund for New York City and Plug and Play.
Sources and notes
Every fact, and where it came from.
- OWASP GenAI Security Project, OWASP Top 10 for Agentic Applications, December 2025.
- TechRepublic, OWASP Drops First AI Agent Risk List, December 2025. Input from more than 100 researchers.
- Microsoft Learn, Agent registry convergence, updated 2026.
- The Register, AWS Agent Registry, April 9, 2026.
- Clutch Security, MCP server deployment research, December 4, 2025. Vendor data, more than 15,000 deployments, modeled for a 10,000 person organization.
- Pluto Security, MCP telemetry, August 2026. Vendor data, four customer environments, sample sizes mostly not stated.
- Cyberhaven Labs, AI adoption on endpoints, 2026. Vendor data.
- Cloud Security Alliance AI Safety Initiative, Shadow AI Apps, May 30, 2026. Source of cited IBM, Reco, Entro and Gartner figures.
- Cloud Security Alliance with Token Security, Autonomous but Not Controlled, April 2026.
- Anomity, AI agent inventory guide. Vendor source, cites Rubrik, Entro Labs, and CSA with Oasis.
- enz.ai, AI inventory guide. Vendor source, secondary.
- Simon Willison, The lethal trifecta for AI agents, June 16, 2025.
- NIST, AI Risk Management Framework, GOVERN 1.6, and the NIST to ISO/IEC 42001 crosswalk. Paraphrased.
- Data Protection Report, EU AI Act timeline update, July 2026.
- Alatirok, How Many AI Agents Per Enterprise in 2026, citing IBM Think 2026.
Run the sprint.
Count first. Then name. Then repeat.
Pick a sponsor, book ten days and write the definition on day one. You will have a first count, a name beside every agent you found and a date for the next one. That is a governed estate, at version one.
Your next step
Take the playbook. Run the sprint. Leave with your number.
Thirty minutes, no slides. Walk through your twelve doors, pick four lenses and set the date for day one.
01
Scope it
We agree the definition and the sponsor.
02
Look
We choose four lenses and plan the read-only pulls.
03
Name it
We plan the owner ladder and the amnesty window.