The Scope Creep Automated reviews of new AI-related working group submissions

Review · August 20, 2026

PACT - Private Agent Consent and Trust Profile

draft-valverde-oauth-pact -00View on Datatracker →

Problem Statement

PACT addresses the challenge of deploying an AI agent that simultaneously requires three distinct things: machine-verifiable identity, explicit human consent, and privacy-preserving delegation. Existing OAuth specifications each solve one of these in isolation — CIBA provides backchannel human consent, DPoP sender-constrains tokens, RAR carries structured authorization intent, and Token Exchange rebinds audience and scope. PACT specifies how these compose into a single unified delegation flow for agentic systems.

How the Spec Fits In

PACT is a composition profile rather than an extension of any single RFC. It layers on top of a large number of existing specifications:

  • OAuth 2.1, PKCE, PAR, DPoP — transport and request security
  • CIBA — backchannel human consent channel
  • Token Exchange (RFC 8693) — audience rebinding and claim narrowing
  • RAR (RFC 9396) — structured authorization details
  • Client Attestation — hardware/software attestation of the agent runtime
  • OIDC pairwise subject identifiers — cross-RP correlation prevention

It introduces a substantial number of new concepts on top of these foundations: a three-tier identity model (durable host / ephemeral session / pairwise identifier), a capability registry and grant system, a usage ledger, an Agent-Assertion JWT type, risk-graduated consent routing, and delegation chain tracking with depth limits.

Critically, PACT has a normative dependency on VEIL (draft-valverde-oauth-veil), another individual draft by the same author that has not been adopted by any working group. PACT defers its identity PII handling and ephemeral delivery mechanisms entirely to VEIL.

High-Level Summary

PACT describes the lifecycle of an AI agent acting on behalf of a human across multiple services. The flow is as follows:

  1. Host Registration: The agent software establishes a durable identity (Ed25519 keypair stored on disk) tied to a specific user-client pair. This persists across process restarts.

  2. Session Registration: Each time the agent process starts, it generates a fresh ephemeral Ed25519 keypair that lives only in memory. The session is registered against the durable host via a signed attestation.

  3. Capability Grants: The authorization server maintains a registry of named actions (e.g., purchase, read_profile) with typed constraints (amount limits, allowed values), daily usage counters, and cooldowns. The agent's session is granted access to specific capabilities from this registry.

  4. Human Consent via CIBA: When the agent wants to perform an action, it sends a CIBA request with a signed Agent-Assertion JWT (binding the request cryptographically to a human-visible description). The AS routes the request to the appropriate consent channel — silent approval, push notification, or biometric challenge — based on the sensitivity of the capability.

  5. Token Issuance: An access token is issued carrying delegation claims (agent, task, capabilities, delegation) along with DPoP binding and pairwise pseudonyms for the target RP.

  6. Token Exchange: When the token crosses an audience boundary to a downstream service, the AS mints a narrowed token that strips agent control-plane claims, recomputes pairwise identifiers for the new audience, and narrows the authorization details to the approved subset. The downstream RP only needs to verify a standard DPoP-bound JWT.

A target RP receiving a downstream-exchanged token sees a clean, standard OAuth access token without needing to understand any of the agent capability machinery.

Interoperability and Necessity

PACT has genuine interoperability value in its core premise — specifying how CIBA, Token Exchange, and DPoP compose for agent delegation is not fully specified anywhere today. However, the spec significantly overreaches by specifying implementation internals that do not affect protocol interoperability:

Items that should be removed:

  • File system paths and permissions (~/.zentity/hosts/..., mode 0600/0700) — these are implementation details for one specific platform. A protocol spec must not dictate how a server or client stores its keys internally.
  • Default session TTL values (1800s idle, 86400s max) — these are policy decisions that differ across deployments. If defaults are needed, they belong in a best-practices document, not a normative spec.
  • Usage Ledger enforcement algorithm — the four-step atomic sequence (cooldown check, count check, amount check, record insert) specifies server-side database transaction behavior. This is purely an implementation concern.
  • JWKS fetch hardening parameters (max 3 redirects, response size cap) — again, server-side implementation choices.
  • The ~/.zentity/ path also reveals that this spec was developed with a specific vendor product (Zentity) in mind, which is inappropriate for a general IETF standard.

Interoperability gaps that remain underspecified:

  • The relationship between PACT's capability model and RAR authorization_details objects is unclear. Both carry structured authorization intent; the spec mentions RAR but doesn't specify precisely how capabilities map to or replace authorization_details in practice.
  • There is no client-facing discovery mechanism for what capabilities exist at a server beyond simply GETting the capability list. It's unclear how a client would request a specific capability before knowing the server's vocabulary.
  • The trust tier model (unverified vs attested) is described but the behavioral differences are minimal. What a server must do differently for each tier is not normatively defined in a way that would produce interoperability.

Normative dependency on an unpublished individual draft:

The normative dependency on VEIL (draft-valverde-oauth-veil) — another unreviewed draft by the same author — is a significant process concern. If PACT were to advance, VEIL would need to advance alongside it or be merged in. Depending on an unreviewed companion spec makes it impossible to evaluate PACT's security properties in isolation.

Writing Quality

The draft is clearly written by someone with genuine technical depth — the mechanisms are described with reasonable precision, security and privacy considerations are present and substantive, and the spec makes appropriate use of RFC 2119 language. However, several issues are worth noting:

  • The document reads more like a design document or architecture guide than a protocol specification in places. Sections like the usage ledger enforcement algorithm and the cryptographic continuity chain would be better described at the protocol boundary rather than as implementation steps.
  • The scope is extremely ambitious for a -00 draft from a single new author. The spec spans identity management, consent routing, cryptographic binding, token exchange, delegation chains, and discovery — each of which is typically a standalone specification. This breadth makes it difficult to properly review any single mechanism.
  • References to non-IETF documents (A2A Protocol, AAP) for claim vocabulary alignment introduce informal normative dependencies. Vocabulary alignment with non-IETF documents can be noted informatively, but the claim names and semantics must be self-defined in the spec if they are normative.
  • The capability derivation table in Section 6 (deriving a capability name from CIBA request properties) feels ad hoc — it would need significant expansion to cover real-world use cases, and any gap in the table leaves implementors without guidance.

Author Background

Gustavo Valverde ([email protected]), Zentity, Portugal

  • RFCs Published: 0
  • New to IETF: Yes — no RFC authorship history and no prior working group involvement visible in Datatracker
  • Active Internet-Drafts: 2, both authored alone (draft-valverde-oauth-veil and draft-valverde-oauth-pact)

This is a new author submitting two tightly coupled individual drafts simultaneously, with no track record in the IETF and no co-authors with IETF experience. The two drafts are mutually dependent (PACT normatively depends on VEIL), creating an insular ecosystem that hasn't yet been exposed to working group review. The Zentity affiliation suggests this work may be motivated by a specific product implementation, which can explain the specificity of implementation details in the spec.

Overall Assessment

PACT tackles a real and unresolved problem in the agentic OAuth space — specifically how to compose CIBA, Token Exchange, and DPoP for privacy-preserving human-in-the-loop agent delegation. The core ideas (pairwise agent identifiers, claim narrowing on exchange, risk-graduated consent routing) are worth exploring and align with ongoing concerns in the working group.

However, in its current form the draft is too ambitious in scope, too prescriptive about implementation internals, and carries a normative dependency on a companion draft that is also unreviewed. The most productive path for feedback would be to encourage the author to:

  1. Narrow the scope to the protocol-level composition (how CIBA + Token Exchange + DPoP interact for agent delegation) and strip the implementation-detail sections
  2. Resolve the VEIL dependency — either merge relevant VEIL mechanisms into PACT or clearly scope PACT to require VEIL as a separate prerequisite
  3. Engage the working group on the capability/grant model vs. RAR before formalizing a parallel authorization vocabulary
  4. Seek co-authors with IETF experience to strengthen the process path