Governed Autonomy and CoSAI’s Open Delegation & Identity Standard

An open draft standard that asks who an agent represents, what authority it holds, and which running instance is acting. Those three questions are this doctrine’s first and fourth Laws and Least Agency, answered at the identity and delegation layer with requirements an implementation can fail.

The Open Delegation & Identity Standard (ODIS) is, in the words of its sponsor, “an open community effort, developed through the Coalition for Secure AI (CoSAI), an OASIS Open Project, that aims to establish a common approach for AI agent identity, delegation, and governance.” It is developed under CoSAI Workstream 4, which focuses on secure design patterns for agentic systems. On 28 August 2026 CoSAI held a hybrid working session on the draft, under the Chatham House Rule, that included two live demonstrations; CoSAI’s account of the session was published on 1 September 2026.

The specification itself is public. Its abstract says ODIS “defines an open, vendor-neutral architecture for establishing cryptographic identity, delegation, and governance for autonomous AI agents operating across enterprise trust domains,” and that it augments OAuth 2.0, OpenID Connect, SPIFFE, SCIM and policy-engine infrastructure: “it does not replace them.” It is organized as four pillars, three implementation layers named Passport, Bridge and Router, a table of identified requirements, and three conformance profiles. Its status line reads “Unapproved contributor draft intended for open development,” and its governance section states that the document “has no OASIS or CoSAI approval status.” This page treats it accordingly: as a draft, mapped as it stands.

This page is a mapping, not a review. It maps ODIS’s framing, pillars, layers and a selection of its requirements onto this doctrine’s constructs, records two places where the two differ by design, and notes where the doctrine’s proposed conformance criteria sit alongside it. ODIS does not cite this doctrine, and this mapping does not suggest otherwise.

Note on criterion identifiers. Identifiers such as C1.1 or CX.1 refer to RFC 001 revision 2, a historical proposal preserved as published. They are not v4 requirements. The current requirements are the baseline R01 to R18 and the Assessment Companion. This mapping has not yet been restated against them.

What was read. CoSAI’s post of 1 September 2026 announcing the working session was read in full on 27 October 2026. The ODIS specification (RFCs/ODIS.md in the cosai-oasis/ws4-odis repository, at commit 148dc41 of 8 September 2026, the latest on its main branch when read) was read on 27 October 2026 for its abstract, Sections 1, 2, 3.3, 3.4, 5, 8.1, 8.4, 9 and 10, and the opening of Section 8.2. The rest of Section 3, and Sections 4, 6 and 7, were not read, nor was the candidate implementation in the same repository, the Workstream 4 repository, the session recording or demonstrations, or any OASIS process document. The specification is a draft that CoSAI describes as still evolving. Where this page and ODIS disagree about ODIS, ODIS is right.

The problem as CoSAI frames it

CoSAI’s post states the problem in one sentence: “Answering that means connecting three things: the principal the agent represents, the authority it has been given, and the specific instance of the agent taking the action.” Each of the three has a home in this doctrine. Left column is CoSAI’s wording, verbatim.

CoSAI’s framingWhere it lands in this doctrine
“the principal the agent represents”Law 4, Trust Does Not Travel, and Plane 5, Multi-Agent Trust & Delegation, for the delegation; the reconstructable record at Plane 4, Human Oversight, Audit & Traceability, which for any governed action preserves sufficient, protected evidence of who requested, authorized and performed it
“the authority it has been given”Least Agency, stated under Law 1: an agent “must be granted only the minimum decision scope, tool access, and action authority required to accomplish its mission.” Architecturally, the mission-scoped access boundaries of Plane 1, Agent Identity & Lifecycle
“the specific instance of the agent taking the action”Law 1, Agents Are Identities, Not Tools: an agent is an attributable operational principal with an accountable owner and lifecycle; Plane 1, including its credential lifecycle management for ephemeral credentials

The four pillars

ODIS names “four foundational pillars that must be addressed for secure, enterprise-grade agentic operations.” These are ODIS’s pillars, not this doctrine’s six Framework Pillars; the shared word is a coincidence of vocabulary. Left column is each pillar’s name and definition from the specification, verbatim.

ODIS pillarWhere it lands in this doctrine
Pillar 1, Delegated Principal Identity. “The mechanism by which an authenticated human or service principal’s intent and authorization context are conveyed to an agent acting on that principal’s behalf.”Law 4 and Plane 5, where delegation is a handoff at which trust is established; attribution to the requesting principal at Plane 4
Pillar 2, Agent Code/Package Identity. “The verifiable identity of the software artifact itself, proving what code or package is intended to run and whether it originates from a trusted source.”Plane 1. The nearest criteria proposed in RFC 001 address provenance for what an agent runs on and calls: C1.5 for the models it may execute under, C2.2 for the tools it can invoke. ODIS states the binding for the agent’s own software artifact directly.
Pillar 3, Agent Runtime Instance Identity. “The cryptographic identity of a specific running agent instance, proving that this exact runtime is the one receiving delegated authority.”Law 1; Plane 1; Framework Pillar 1, Agent Identity
Pillar 4, Cascaded Delegation. “The secure propagation of authority, identity context, and intent when one agent invokes another agent across one or more trust boundaries.”Law 4; Plane 5; Framework Pillar 6, Multi-Agent Governance. See below for how the two treat a sub-agent’s authority.

The three layers

ODIS organizes its architecture as three implementation layers, and says each “has a defined role, clear input/output interfaces, and explicit boundaries.” Like this doctrine’s Planes, they are logical rather than physical: the specification says a conformant deployment may place them “in a gateway, sidecar, service mesh, SDK, cloud service, or any combination thereof.”

ODIS layerWhere it lands in this doctrine
Layer 1, The Passport (Identity & Attestation). Short-lived Agent Runtime Credentials issued after attestation and resolved to an Agent Registration Record.Plane 1, Agent Identity & Lifecycle
Layer 2, The Bridge (Delegation & Access). Delegated authorization, attenuation across delegation hops, and translation to downstream systems that do not understand ODIS.Plane 5 for delegation; the mission-scoped access boundaries of Plane 1, where Least Agency is enforced
Layer 3, The Router (Discovery & Governance). In the specification’s words, it “Defines WHERE the agent can go, what it can do there, and which policy context governs each action.”Plane 2, Execution & Tool Governance, for tool and API invocation authorization; Plane 3, Policy & Compliance Engine, for the policy decision; Law 2, Enforce at Runtime

Selected requirements

ODIS states its normative content as identified requirements, each tagged with the conformance profile it belongs to. The ones below have a direct counterpart in this doctrine. Left column quotes the requirement, verbatim, in part.

ODIS requirementWhere it lands in this doctrine
ODIS-L2-01, Delegated Authorization (Core). The system “MUST compute the effective authority as the intersection of the originating principal’s current authority, the active Agent Registration Record, any parent Delegation Record, the requested task and resource, applicable environmental constraints, and the Provider Adapter mapping. If that intersection cannot be affirmatively established, the request MUST fail closed.”Least Agency under Law 1; the fail-closed rule is RFC 001’s C3.8: an action whose applicable criteria cannot be evaluated does not proceed
ODIS-L2-06, Authorization Attenuation (Core). “When an agent delegates to a sub-agent, the sub-agent MUST receive authority equal to or narrower than its parent over resource, action, and constraint dimensions.”Law 4; Plane 5, where authority is established at each handoff and the recipient is authorized at its own boundary. See the next section.
ODIS-L2-13, Presenter Authority Scoping (Core). “The model-visible path MUST NOT be given a generic arbitrary-payload signing or proof-construction interface for holder-bound downstream credentials.”RFC 001’s C1.4: no path within the agent’s enumerated tool surface reaches its own authority, mission scope, or governing policy; and C3.1: the rules governing an agent’s behavior are external to the agent
ODIS-CC-05, Governed Identity Creation (Core). “An agent MUST NOT approve, create, or expand its own registration or mint credentials for another identity unless a separately authorized provisioning role permits it.”Law 1 and Plane 1’s agent registry; RFC 001’s C1.4
ODIS-L2-02, Bounded Authorization (Core). “When human approval is required, the approval MUST be bound to the Agent Registration Record, task, requested authority, resource audience, constraints, and expiry.”Law 5, Humans Retain the Right to Intervene; Framework Pillar 5, Human Oversight & Intervention; RFC 001’s C4.3 and C4.4
ODIS-L3-04, Revocation Latency and ODIS-L3-05, Kill Switch (Core). Revocation events within a trust domain must propagate within the declared maximum latency, “not exceeding 300 seconds from receipt at the domain boundary,” and “The system MUST support immediate global de-provisioning of an agent identity via a single operation.”Law 5; trust revocation propagation at Plane 5; RFC 001’s C4.1: a revocation, override or halt by an external authority is honored before the agent’s next commit
ODIS-L3-03, Velocity Limits (Core). “The system MUST enforce configurable rate limits per agent, especially for destructive operations.”Law 2; action-level decision interception at Plane 2
ODIS-L3-07, Task-Bound Tokens (Extended). “Tokens issued to agents SHOULD carry a declared task purpose.” ODIS lists purpose drift in its threat model as an agent that “reuses authority obtained for one purpose to perform unrelated actions.”Framework Pillar 2, Mission Definition; RFC 001’s C1.3: an agent’s authority is granted per mission, expires with the mission, and does not persist across missions
ODIS-CC-02, Dual-Identity Audit Trail (Core). “Every logged action MUST identify the logical agent, the executing runtime instance, and the authenticated originating principal.”Plane 4; RFC 001’s C1.1 (every action attributed to exactly one agent, in every system it touches) and C1.2 (agent-originated actions distinguishable from human-originated actions without inference)
ODIS-L1-10, Accountable Sponsor (Core). “Every agent identity MUST have an accountable human sponsor or named service-owner record.”Plane 1’s identity lifecycle; Law 5

Where the two differ by design

Scope. ODIS draws its own boundary plainly: “ODIS is an identity and delegation standard. It does not standardize organization-specific policy logic or a policy language.” Its limitations section adds that it does not address “Model-level safety or hallucination,” nor prompt-injection prevention as a content-analysis problem. This doctrine’s scope includes the behavioral layer as well: Framework Pillar 3, Behavioral Policy, the policy engine at Plane 3, and the four threats on its Threat Surface. The two fit together rather than overlap. ODIS specifies a structured identity-context object that external policy engines consume (ODIS-L3-06), which is the kind of input Plane 3 evaluates.

A sub-agent’s authority. ODIS derives a sub-agent’s authority from its parent’s by attenuation: equal or narrower, computed as an intersection that also includes the sub-agent’s own active registration record (ODIS-L2-01, L2-06). This doctrine’s Law 4 holds that a handoff carries a task and authenticated context, not automatic authority, and that delegation shall not amplify what the delegator is authorized to grant. RFC 001’s historical C5.2 stated that “Authority is never inherited from an invoking agent,” and that a subagent’s authority is “granted independently, scoped to its own mission, by the same external authority that grants the invoker’s.” Both rule out privilege amplification down a chain; they describe its mechanism differently. Whether an attenuated grant bounded by an independent registration record satisfies the v4 delegation requirement (R08) is a question for assessment against a declared profile, not one this page settles.

What this doctrine provides alongside it

ODIS is not a list of questions. It already states normative requirements in BCP 14 language and defines three conformance profiles, each a strict superset of the one before: Core, Extended and Safety. A profile claim, it says, applies to “a named, versioned deployable system boundary,” and the draft includes a conformance-declaration template. For the identity and delegation layer, ODIS is the more detailed of the two documents.

The doctrine’s contribution sits at a different altitude. RFC 001, The Conformance Layer proposes twenty-nine criteria for what must be true for a single agent action to count as governed, across all five Planes rather than at the identity and delegation boundary alone. It is a historical proposal, not doctrine; the doctrine in force is v4. The criteria that reach past ODIS’s stated scope include:

ConcernRFC 001 criterion (proposed)
Grading actions by consequence before they are attemptedC4.2: every governed action carries an assigned tier, determined before it is attempted, recorded with the reversibility basis, recovery window and magnitude that produced it
What a human approver actually sawC4.4: the representation of an action presented to a human for authorization is generated by the governing system from the action’s committed parameters, not by the agent, and the record retains exactly what was shown
Why a policy decision came out as it didC3.3: the basis of every policy decision is recorded at decision time, traceable to a specific rule. C3.4: given the recorded decision inputs and the policy version in force, a past decision reproduces exactly
Effects no rule addressesC3.5: an action whose effect is not enumerated is denied, not permitted
Stopping what an affected agent already producedC5.4: on detection of anomalous behavior, the governing system halts every downstream consumer of the affected agent’s output, not only the agent
All of it holding at onceCX.1, the Simultaneity Criterion: every applicable criterion holds for a single action within that action’s decision window, and the parameters evaluated are identical to the parameters committed

An identity and delegation standard and a conformance layer answer different questions. The first says how an agent proves who it is, whom it acts for and what it may do; the second says what must be true of a given action for it to count as governed. They are designed to be used together: an ODIS deployment supplies much of the identity and delegation evidence that RFC 001’s Plane 1 and Plane 5 criteria ask for. CoSAI invites review of the ODIS draft on GitHub and through Workstream 4; that invitation is CoSAI’s, and nothing on this page responds to it. The same delegation question is mapped on this site against NIST’s agent identity and authorization work and, for provisioning, the SCIM Agent Governance Extension.

Primary sources

Mapped 27 October 2026 against the ODIS draft at commit 148dc41. ODIS is an unapproved draft; when it advances or changes, this page will be revised to match it rather than defended.