Review · August 26, 2026
SCIM Agent Governance Extension
draft-kushwaha-scim-agent-governance-00 — Saurabh Kushwaha (Oracle), submitted 26 July 2026. SCIM working group, intended status Informational.
The Problem Statement
The Agent resource in draft-wzdk-scim-agent-resource has one field for whether the agent is on or off: a boolean active. That's not enough to run an agent fleet.
This draft adds four things on top of it:
- A five-value lifecycle state instead of the boolean, so you can tell "provisioned but not turned on" from "temporarily disabled" from "frozen for an investigation" from "permanently dead"
- A human-oversight classification (
autonomyLevel) - A validity window (
validFrom/validTo) - A pointer to where the agent's credential binding methods are published
The lifecycle values come from ISO/IEC 24760-1 rather than being invented, with one addition: quarantined.
The genuinely interesting part isn't any of the four attributes. It's Section 5, "Out of Scope," which lists four attributes that were proposed and cut, with the reasoning for each. More on that below, because it's the best thing in the document.
How the Spec Fits In
This is a SCIM extension schema, the same shape as the Enterprise User extension in RFC 7643 §4.3 or the device extensions in RFC 9944. One new URN, five new attributes, one IANA registration. No new endpoints, no new protocol operations, no new roles.
Dependency count is small: RFC 7643, RFC 7644, RFC 3339, RFC 8259, plus the base Agent resource draft. That's it for normative references beyond BCP 14.
Two structural notes matter more than the count.
It normatively depends on an unadopted individual draft. The base Agent resource draft is at -00 and is not a working group document. If the WG changes the Agent design during adoption, this extension changes with it. The acknowledgments credit Danny Zollner and Pamela Dingle, two of the base draft's authors, for shaping the scope, so this is a friendly follow-on rather than a competing proposal. But the sequencing question is real: is this a separate document, or is it Section 5 of the base draft?
It redefines a core-schema attribute. §3.1.5 turns active from a client-settable boolean into a value the provider MUST derive from lifecycleState.state. An extension schema changing the mutability semantics of the base schema's required attribute is unusual, and it's where most of the interop problems in this draft come from.
The relationship to other adopted SCIM work is mostly right but has one gap. §3.1.6 correctly says that lifecycle transitions are point-in-time security events that SCIM polling can't deliver promptly, and routes them to the Shared Signals Framework. Good instinct. But it cites the generic OpenID SSF spec and never mentions RFC 9967, the SCIM WG's own SET profile published in May 2026, which already defines urn:ietf:params:scim:event:prov:activate, prov:deactivate, and prov:patch. That's the mechanism this section is asking for, and it's already an RFC from this working group.
Against the charter, the hook is clean. "Updates to account state for capturing context of the state or change in state of the users account" is an explicit charter item for the RFC 7643/7644 revision. This draft is that work, applied to agents.
High-Level Summary
Today a provisioned agent is either enabled or disabled. That single bit has to carry every operational situation an enterprise actually encounters, and it can't.
Consider what happens when a security team detects that an agent has been compromised. They disable it. Now the record says active: false, which is indistinguishable from "we haven't finished setting it up yet," "we paused it for the quarter," and "we decommissioned it last year." The investigation team has no way to query for "agents currently under investigation," and nothing stops an administrator from editing the record while it's being investigated, destroying the evidence.
This draft replaces the bit with a state machine. Five states, an enforced transition graph, and a timestamp plus actor recorded on every transition.
For an access-review process, the payoff is queryability. You can ask the provisioning system for every agent frozen pending investigation, or every agent whose authorization window expired last month but is still running. Those are the two example queries in the draft, and they're the right examples. Both are things auditors ask for and nobody can currently answer.
For the security team, quarantined is a state with teeth. Suspension is an administrative convenience whose enforcement the draft leaves to the deployment. Quarantine says all access MUST be blocked immediately, tokens MUST be revoked, and the record MUST NOT be modified except by the review process. The reasoning is specific to agents and it's correct: a compromised agent keeps acting at machine speed while humans are still reading the alert, and investigating it depends on freezing its relationship graph as of the detection moment.
For decommissioning, archived is deliberately terminal. You cannot un-archive an agent. The draft breaks from ISO 24760-1 and from OpenID Provider Commands on this point, on the grounds that restoring an identity whose authority was withdrawn for cause is a quiet privilege-escalation path, and that reviving an identity breaks the link between it and its audit history. The intended path is to provision a replacement.
The catch a leader should understand: that last decision has an unhandled consequence. The base draft requires agentUserName to be unique across every Agent resource in the domain, and this draft says archived records SHOULD be retained rather than deleted. Put those together and every archive permanently burns the agent's name. Archive invoice-review-agent by mistake and you can never provision invoice-review-agent again. Terminal states plus name uniqueness plus retention is a combination that will surface in the first month of any real deployment, and the draft doesn't address it.
The second catch is layering. This is a provisioning schema, but a good number of its normative requirements are about things provisioning systems don't do: revoking tokens, blocking access, declining to issue credentials. In most enterprise deployments the SCIM service provider is the target application, not the authorization server. A schema document can't impose testable requirements on a component it doesn't talk to. The requirements themselves are correct as deployment guidance. They just can't be conformance criteria.
Significant Interoperability Concerns
The URN inverts the SCIM naming convention. Every registered SCIM extension URN ends with the resource type it extends: extension:enterprise:2.0:User, extension:ble:2.0:Device, extension:groupMembers:2.0:Group. This draft uses extension:agent:2.0:Governance, putting the resource type in the extension-name slot and a concept in the resource-type slot. It should be urn:ietf:params:scim:schemas:extension:governance:2.0:Agent. The author's own DID/VC draft uses extension:didvc:2.0:User correctly, so this looks like an oversight rather than a considered choice. Easy fix, but it needs to happen before the URN gets registered and copied into implementations.
active and the validity window contradict each other. §3.1.5 says active MUST be derived from lifecycleState.state. §3.3 says a provider SHOULD treat an out-of-window agent as suspended, and explicitly says the stored state value is not modified by the passage of time. So the day after validTo passes, the state is still active, the derived active is still true, and the provider is treating the agent as suspended. Anything reading active gets the wrong answer. §3.1.7's second example filter looks for exactly this condition and calls it "not reconciled," which shows the author is aware of the drift, but §3.1.5's MUST makes active actively misleading rather than merely stale. Two implementations will make opposite choices about whether to flip the bit.
Attaching the extension to an existing Agent is undefined. §3.1.1 says the state SHOULD default to established on creation, and established requires active: false. What happens when a provider adds this extension to an Agent that already exists with active: true? Nothing in the draft says. This is the most common real-world path, since deployments add extensions to existing fleets.
PUT collides with the transition table. §3.1.4 says the provider MUST reject any modification whose resulting transition is not in Table 2. SCIM PUT replaces the whole resource, so every PUT carries lifecycleState.state, and a PUT that doesn't change the state produces a same-state "transition" that isn't in the table. Read literally, every idempotent PUT gets a 400. The draft needs an explicit statement that a request leaving the state unchanged is not a transition and does not update transitionedAt.
credentialDiscoveryUri has no interoperability content. It's a URI pointing at a document whose format is explicitly out of scope. Two conformant implementations can populate this attribute and be unable to consume each other's documents. It also carries a requirement that a provider "MUST validate that the value refers to an authority authorized for the agent or its operator" with no defined mechanism for doing so. This attribute is a placeholder for future work, and it should be labeled as one or held back.
reason is free text carrying a distinction the design depends on. §3.1.2 collapses revoked and decommissioned into a single archived state and says the difference is carried in reason. Then §3.1 says reason codes are not standardized and deployments SHOULD document their own vocabulary. So the audit distinction the draft claims to preserve doesn't survive crossing a domain boundary, which is the situation SCIM exists to handle. Either register a small canonical set for the archival reasons, or say plainly that the distinction is local.
autonomyLevel boundaries won't classify consistently. semiAutonomous is defined as requiring "human initiation or checkpoint approval," which are two materially different arrangements folded into one value. And the line between autonomous and semiAutonomous turns on whether a human started the task, so the same agent is classified differently depending on whether it's triggered by a person clicking a button or by a scheduled job. The draft names the EU AI Act as an anticipated consumer, which raises the stakes on getting the boundaries crisp. Right now three deployments will produce three different classifications of the same agent.
The IANA section is incomplete. RFC 7643 §10.3.2 defines the schema registration template: Schema URI, Schema Name, Intended or Associated Resource Type, Purpose, Single-value Attributes, Multi-valued Attributes. §8 supplies three of six fields. Worth noting alongside this: RFC 7643 §10.3.1 allows URNs outside the core and api namespaces to be registered first-come-first-served with a simple review by the designated expert. This extension does not require an RFC to exist. The argument for making it one is that the working group wants a common extension rather than a vendor-specific one, and that the value sets are meant to be referenced from other protocols. That's a legitimate argument and RFC 9944 is precedent, but the WG should make the choice deliberately rather than by default.
Mechanisms That Could Be Removed
The document is 21 pages to define five attributes, roughly twice the length of the base draft it extends. Most of the excess is justification rather than mechanism, which is the better failure mode, but it should still come down.
Figure 1, the ASCII transition graph. It duplicates Table 2, and it's drawn wrong: suspended appears twice, and the arrows into quarantined don't correspond to the table. Table 2 is clear on its own. Delete the figure.
The direct quarantined to active transition. §3.1.4 permits it, then §3.1.5 says implementations SHOULD go through suspended first, and both places hedge it with policy restrictions and administrative-approval language. Just require quarantined to suspended to active. That removes a table row, two paragraphs of qualification, and the notion of "restricted administrative principals," which the draft can't define anyway.
§5, Out of Scope, should shrink to an appendix. This is a strange recommendation because §5 is the most valuable section in the draft for the working group right now. It records four attributes that were proposed and cut, with the reasoning, and it routes each concern to the right venue: risk posture to SSF, delegation scope to RFC 8693 and WIMSE, tool permissions to the authorization layer. That's genuinely useful review material and it's evidence the author listened to the list. It is not specification, and in a published RFC it will read as an argument nobody can follow the other half of. Keep it for now, move it to an appendix before publication.
§3.1.3 can compress. A page of prose justifying one new canonical value is more than the value needs. Two sentences carry the argument: a compromised agent keeps acting between detection and response, and investigation requires the record be frozen. The rest is context for the list.
There's also an unstated design rule worth making explicit, because a reviewer will otherwise read the draft as inconsistent. quarantined gets its own state while revoked and decommissioned get folded into reason. The rule the author is applying appears to be: distinctions that change what the provider must enforce become states, and distinctions that only change what the audit log says become reason codes. That's a good rule. Say it in §3.1.2 and both decisions become obvious instead of arbitrary.
Writing Quality
This is a well-written -00, and that's worth saying up front because it's not the norm.
BCP 14 keywords are used correctly and sparingly. Sentences are well formed. Terminology is defined before use and inherited explicitly from the base draft rather than redefined. References are complete, with titles, authors, and dates, and the normative and informative split is present and correctly allocated. Tables are properly formatted with captions. The example is realistic. Every non-obvious decision has a stated reason, and where the draft diverges from a normative source it says so and explains why, which is exactly what §3.1.2's treatment of the ISO restore path does.
The defects are minor and mechanical:
- §3.1 says transitions are subject to Section 3.1.5. The transition rules are in 3.1.4.
- Figure 1 is garbled, as described above.
- "To the author's knowledge no ISO/IEC 24760-1 equivalent for agent identities exists" is informal for an RFC. Drop the first clause.
- There's no "About This Document" block pointing at the SCIM mailing list. The base draft has one.
- Intended status is Informational, but the document contains enforcement MUSTs and requests an IANA registration. The base Agent draft is also Informational, so this is consistent with its parent, but RFC 9944 published device extensions as Standards Track. Worth asking the chairs which track the agent work is on before the extension picks one.
On concision: the ideas fit in about 12 pages. The excess is in §3.1.3 and §5, both of which are justification aimed at the working group rather than at implementers. That's the right thing to over-invest in at -00 and the wrong thing to publish.
Background of the Author
Saurabh Kushwaha (Oracle Corporation) — 0 RFCs, 4 active Internet-Drafts, 0 expired. No IETF group roles listed. This is a new IETF author, and a prolific one right now.
All four active drafts are SCIM:
draft-kushwaha-scim-agent-governance(this one)draft-kushwaha-scim-didvc-binding(at -01, so it's being revised)draft-kushwaha-scim-tenant-resourcedraft-kushwaha-scim-attr-cursor-pagination
Four individual drafts into one working group from a participant with no RFCs is a lot of surface for the WG to absorb. Set against that, the topics are well chosen. Cursor-based pagination for multi-valued attributes sits on an explicit charter item and complements RFC 9865 rather than duplicating it, and tenant-aware provisioning is a real gap. This isn't scattershot.
The acknowledgments are the most informative thing about the author here. They credit Zollner and Dingle by name for the three specific pieces of feedback that shaped the draft: ground the lifecycle values in ISO 24760-1, use a discovery document instead of embedded credential metadata, and route the real-time and cross-domain concerns to SSF and WIMSE. All three are visible in the document, and §5 is the record of what got cut as a result. This author engaged on the list, got told to narrow the scope, and narrowed the scope. That is not common, and it's the strongest signal in the package.
Note the affiliation is Oracle but the address is a personal Gmail account. Common enough, and not a substantive issue, but it means the draft doesn't carry a company's position on it.
Suggested Feedback
If you engage on the list, this one deserves a materially warmer opening than most -00 submissions:
Lead with §5. Say plainly that enumerating the cut attributes with reasoning, and pointing each one at the right venue, is how this should be done, and that you'd like to see other agent-related drafts adopt the pattern. It costs nothing and it reinforces exactly the behavior you want from new participants.
Ask the sequencing question. Should this be a separate document, or a section in the base Agent draft? The base draft is at -00 and unadopted, and building an extension on it now means both documents move together. Zollner and Dingle already shaped this scope, so the answer probably belongs to them, but it's better asked before the WG has two documents to adopt.
The URN is backwards.
extension:governance:2.0:Agent, notextension:agent:2.0:Governance. Concrete, cheap, and needs to land before anyone registers or implements it.activecannot be both derived and stale. The interaction between §3.1.5's MUST-derive and §3.3's out-of-window handling producesactive: truefor an agent the provider is treating as suspended. This is the most consequential technical issue in the draft because it makes an existing, widely-consumed attribute lie.Terminal
archivedplusagentUserNameuniqueness burns the name permanently. Raise this concretely, because it's an operational problem that will hit the first deployment and the fix affects the design of the terminal state, not just the text.Point at RFC 9967. §3.1.6 is asking for the SCIM SET profile and doesn't cite it. The
prov:activateandprov:deactivateevents are also keyed on theactiveboolean that this draft redefines, so there's an interaction to work out rather than just a reference to add.Raise the layering question about token revocation. The requirements are right as deployment guidance and wrong as conformance criteria, because the SCIM service provider frequently isn't the token issuer. Suggest moving them out of normative language and into security considerations framed as guidance to deployments.
On the merits: this is a small, well-scoped extension addressing a real gap, on a charter item, from an author who has demonstrably absorbed working group feedback. The lifecycle state model is the part worth keeping and it's mostly right. credentialDiscoveryUri is a placeholder that should wait for a document format. autonomyLevel needs sharper value definitions before anyone builds regulatory reporting on it. None of the problems above are structural, which puts this in better shape at -00 than most of what's arriving in the agent identity space right now.