A developer stands up an AI agent with live cloud keys between coffee and lunch. A coding assistant holds full access to a source tree and feeds customer data into its prompts for months, surfaced only because someone mentions it in passing. This is the normal state in most enterprises right now, not the exception.
Advisory
Agentic AI Identity
AI agents are landing across your business faster than identity, security, and governance can see them. We give your team the discipline to see every agent, govern what it can do, and shut it down when you need to.
The Control Question
You cannot govern what you cannot see
The question is no longer whether AI is in your building. It is whether you can see it, govern it, and stop it when you need to.
The Vision · Lens 1
An agent is a joiner nobody onboarded
You already run a joiner-mover-leaver lifecycle for every employee. An agent is a new kind of joiner: one nobody onboarded, nobody owns, and that has no leaver process to switch off. The discipline is familiar. The machinery you run it with cannot see this new joiner.
See it
Find the agents that skipped the front door, and give each one an owner.
Why your tooling cannot
Your IGA never onboarded it. It authenticated past your controls without ever being registered, so it is unattributable: no owner, no record, no review.
Govern it
Permit it precisely, and prove it only ever acts in scope.
Why your tooling cannot
It signs in with a key in an environment variable, not anything your IdP issued. An access review here means reading its behaviour, not clicking "yes, still needed".
Kill it
Revoke it everywhere at once, under incident pressure.
Why your tooling cannot
Universal Logout only kills agents your IdP minted. The rest live in the cloud, in SaaS, in tool connections, and need rotation you have no runbook for.
The ten stages underneath, from the decision to build to the day you kill it
01Intent02Discovery03Registration04Authentication05Authorize as itself06Authorize on behalf of07Telemetry08Drift09Certification10Decommissioning
The Vision · Lens 2
No single tool covers it. Identity is the spine.
Governing agents is not one product. It lands across six layers, and identity is the through-line that runs across all of them: the one place an agent is a registered, owned, revocable thing.
EdgeNetwork and secure-access visibility. The early warning that an agent is reaching somewhere new.
InteractionThe browser and endpoint, often the only place to see an actual prompt, upload, or output.
APITokens and gateways. How AI is wired into your estate.
RuntimeThe unsettled layer between the user, the model, the prompt, and policy.
IdentityWho and what. Lifecycle, privilege, delegation. The spine the rest hangs on.The through-line
SOCDetect and respond. Behavioural analysis and incident response.
No single vendor owns the whole picture. That gap is the white space we work in. You cannot buy your way out of it. You need architecture.
The Approach
How we advise
We do not sell a tool, a platform, or a managed service. We sell judgment and architecture, delivered as an advisory engagement in four movements. The deliverable is an architecture your own team runs, not a dependency on us.
01
Take a position
Decide what kind of AI adopter you are, on the record, at executive level. Every control downstream is cheaper once the operating model is set.
02
See reality
Detection in parallel with policy. We surface the agents that skipped the front door through SaaS, the endpoint, and the cloud. Reality validated by telemetry, not a survey.
03
Design controls
Who registers the agent, how it proves it is the agent, what it may do as itself and on a user's behalf, and how every one of those is enforced and logged. We lead from identity.
04
Keep it assured
Telemetry into your SOC, drift detection when a model or owner changes, recertification that reads behaviour, and a decommissioning path that revokes everything at once.
The Approach · Across Your Stack
We place the platforms. We do not push them.
We design vendor-neutrally and lead from identity. In practice a few platforms carry most of the load, each owning a different part of the lifecycle.
Identity backbone
Okta
Who this agent is and what it is allowed to be. Register, authenticate, authorize as itself, log, recertify, revoke.
Delegation and consent
Auth0
The agent acts as a specific user, lawfully and auditably. Consent, step-up, the user genuinely in scope, never a broad service account.
Point of interaction
Island
Identity decides the policy; the browser governs the work that identity does. It sees the prompt an encrypted connection hides, and can cut a session mid-incident.
The Approach · Who This Is For
For the people who already own identity
The CISO, the head of IAM, and the identity or platform architect who own the human-identity program and now watch an AI surface expanding underneath them with no clear owner. Usually triggered by a board question, a near-miss, or a regulator.
This week
Turn on the visibility you already pay for.
This month
Take a position at executive level.
This quarter
Enforce it, and extend deliberately.
The standards for agent identity are still being written, with competing drafts and no convergence yet. We counsel informed patience: activate what you own, wrap only the genuinely new layer, and buy that part short and portable. Watch the drafts, not the marketing.
Where This Connects
Where this connects
Identity by Design
The architecture discipline underneath. Agent identity is one surface of the control plane we design.
Workforce Identity
The backbone that registers, scopes, and revokes. Where an agent becomes an owned, governable identity.
IAM Assessment
Find the agents first. Discovery of unowned agents and their scope is where most engagements start.
Identity Resilience
Roll back what a rogue agent changed. Point-in-time recovery for the identity control plane.
What should we do about AI agents?
That is the question we hear most, and it is exactly where we start. Bring us the question your board is already asking, and we turn it into a position, a map of your agent estate, and an architecture your own team can run.
Start the ConversationReference · The Full Map
The complete model
The whole framework in one grid, for teams that want the depth and for machines that cite it. You do not need to read this to work with us. The page above is the argument; this is the appendix.
| Lifecycle stage | Identity | Edge | Interaction | API | Runtime | SOC |
|---|---|---|---|---|---|---|
| 01 Intent | ||||||
| 02 Discovery | ||||||
| 03 Registration | ||||||
| 04 Authentication | ||||||
| 05 Authz as itself | ||||||
| 06 Authz on behalf | ||||||
| 07 Telemetry | ||||||
| 08 Drift | ||||||
| 09 Certification | ||||||
| 10 Decommissioning |
Note the Identity column: the densest in the grid. The page's whole thesis, stated by the data rather than asserted. Rows 01 (Intent) and 08 (Drift) sit almost empty: the two controls no vendor owns, which is the architecture we do.