# Delegated Trust Layer Specification (v0.1 Draft)
## Overview
This document defines a draft technical specification for a pluggable trust layer for AI agents. The design is intended to work with agent interoperability protocols such as A2A and MCP while relying on established identity and authorization patterns such as workload identity, OAuth-style delegation, and runtime policy enforcement.[1][2][3]
The specification focuses on five technical pillars: identity, delegation, discovery, enforcement, and auditability.[4][5][6]
## Design principles
- Delegation over impersonation: an agent keeps its own identity while acting on behalf of another principal.[7][8]
- Protocol agnosticism: the trust layer should wrap MCP, A2A, HTTP, gRPC, and other transports without redefining them.[2][9]
- Pluggability: identity providers, policy engines, and audit sinks should be replaceable.[3][10]
- Progressive adoption: existing OAuth, OIDC, IAM, API key, and workload identity systems should be mappable into the trust model.[11][12][13]
- Continuous verification: identity, delegation, and policy should be checked at each hop instead of only once at session start.[5][14]
## Core entities
| Principal | A human user, service, organization, or system that owns authority. [15] |
| Agent | A software actor with its own identity that can perform actions directly or on behalf of a principal. [7][8] |
| Delegation | A bounded grant of authority from a principal to an agent. [4][16] |
| Tool | A callable service, API, or resource that an agent wants to access. [2][17] |
| Trust Domain | An administrative boundary in which identities, policies, and issuers are recognized. [10][18] |
| Policy Decision Point | The component that evaluates whether a requested action is allowed. [19][20] |
| Enforcement Point | The component that blocks or permits requests inline. [5][21] |
## Architectural components
| Agent Naming Service | Assigns stable agent names and binds them to identity records. [6] |
| Agent Registry | Stores signed identity metadata, capabilities, supported protocols, and endpoint hints. [6][22] |
| Identity Provider Bridge | Resolves and validates agent identity through SPIFFE, OIDC, cloud workload identity, or custom PKI. [23][10][18] |
| Delegation Service | Issues, signs, validates, and revokes delegation tokens. [15][4] |
| Resolution Service | Maps trusted agent names to endpoints, gateways, or service descriptors. [6] |
| Protocol Adapter Layer | Normalizes MCP, A2A, HTTP, and gRPC requests into a standard authorization envelope. [2][9] |
| Policy Engine | Evaluates action, scope, resource, time, risk, and intent. [19][24] |
| Runtime Gateway | Enforces identity and policy at request time. [5][21] |
| Audit Log | Persists request chain, decisions, approvals, and outcomes. [25][16] |
| Control Plane | Exposes management APIs for policy, approvals, simulation, and reporting. [26][27] |
## Logical flow
```mermaid
sequenceDiagram
autonumber
participant U as User / Principal
participant DS as Delegation Service
participant A as Agent
participant RG as Runtime Gateway
participant PE as Policy Engine
participant T as Tool or Downstream Agent
participant AL as Audit Log
U->>DS: Request bounded delegation
DS-->>U: Signed delegation token
U->>A: Task + delegation token
A->>RG: Request with agent identity + token
RG->>PE: Evaluate action, scope, context
PE-->>RG: Allow / deny + reason
RG->>T: Forward authorized request
T-->>RG: Result
RG->>AL: Write audit event
RG-->>A: Response
```
GitHub-flavored Markdown supports Mermaid code blocks, which makes this format suitable for technical docs stored in repositories.[28][29]
## Identity document schema
The identity document should describe who the agent is, how it proves identity, and how other systems can discover and trust it.[30][6]
### JSON schema goals
The schema should be versioned, explicit, and easy to validate in tooling. OpenAPI 3.1 aligns with modern JSON Schema support and is a practical base for future API and schema publication.[31][32]
### Example identity document
```json
{
"spec_version": "0.1",
"kind": "AgentIdentityDocument",
"agent_id": "agent:example:scheduler:v1",
"display_name": "example Scheduler Agent",
"owner_id": "org:example",
"issuer": "https://trust.example.ai",
"identity_type": "spiffe",
"subject": "spiffe://example.ai/agents/scheduler",
"public_keys": [
{
"kid": "key-2026-01",
"kty": "OKP",
"crv": "Ed25519",
"x": "base64url-public-key"
}
],
"supported_protocols": ["a2a", "mcp", "https"],
"supported_auth_methods": ["delegation_token", "oidc", "mtls"],
"capabilities": [
"schedule_meeting",
"send_follow_up_email",
"read_calendar_availability"
],
"endpoints": [
{
"protocol": "a2a",
"url": "https://agents.example.ai/scheduler/a2a"
},
{
"protocol": "mcp",
"url": "https://agents.example.ai/scheduler/mcp"
}
],
"attestation": {
"type": "workload",
"issuer": "spire://example-prod",
"evidence_ref": "urn:attest:spire:cluster-a"
},
"created_at": "2026-06-01T20:00:00Z",
"expires_at": "2026-06-08T20:00:00Z",
"signature": "base64url-signature"
}
```
### Required fields
| `spec_version` | Enables versioned parsing and migration. [31][32] |
| `kind` | Distinguishes the object type. |
| `agent_id` | Stable canonical identifier. [6] |
| `owner_id` | Binds the agent to an owning tenant or principal. [15] |
| `issuer` | Identifies the authority publishing the identity document. [33] |
| `identity_type` | Indicates the proof system, such as SPIFFE or OIDC. [10][3] |
| `subject` | Carries the provider-specific subject or workload identifier. [10][18] |
| `public_keys` | Supports signature verification and rotation. [33] |
| `supported_protocols` | Supports protocol negotiation. [2][1] |
| `supported_auth_methods` | Declares accepted trust mechanisms. |
| `endpoints` | Supports routing and resolution. [6] |
| `signature` | Protects document integrity. |
## Delegation token schema
The delegation token is the artifact that binds an agent to bounded authority from a delegator.[4][16]
### Example delegation token payload
```json
{
"spec_version": "0.1",
"kind": "DelegationToken",
"token_id": "dlg_01J0EXAMPLE",
"issuer": "https://trust.example.ai",
"agent_id": "agent:example:scheduler:v1",
"delegator_id": "user:jake-abendroth",
"owner_id": "org:example",
"audience": [
"tool:google-calendar",
"tool:gmail"
],
"intent": "schedule_demo_and_send_confirmation",
"allowed_actions": [
"calendar.create_event",
"calendar.read_availability",
"gmail.send_message"
],
"resource_constraints": {
"calendar_ids": ["primary"],
"email_domain_allowlist": ["example.com"]
},
"max_spend": {
"amount": 0,
"currency": "USD"
},
"max_delegation_depth": 0,
"approval_policy": {
"mode": "required_for_external_email"
},
"issued_at": "2026-06-01T20:10:00Z",
"expires_at": "2026-06-01T20:40:00Z",
"nonce": "random-nonce",
"signature": "base64url-signature"
}
```
### Required semantics
- Tokens should be short-lived and revocable.[34][15]
- The token should bind both the agent and the delegator, not just one of them.[4][7]
- The token should specify explicit allowed actions rather than broad trust statements.[24][4]
- The token should support intent-bound delegation so permissions remain tied to purpose.[16]
- Downstream redelegation should be disabled by default and enabled only with explicit depth controls.[4]
## Request verification algorithm
A receiving trust gateway should evaluate requests in the following order:
1. Normalize the incoming request through a protocol adapter.[2][9]
2. Resolve the presented agent identity from the identity document or registry.[6][30]
3. Verify the identity signature, key validity, and expiration.[33][18]
4. Validate the delegation token signature, issuer, audience, and lifetime.[4][35]
5. Confirm the token binds to the presented `agent_id`.[7][8]
6. Evaluate requested action against `allowed_actions`, resource constraints, local policy, and runtime context.[19][24]
7. If required, invoke approval workflow or additional trust checks.[26][27]
8. Execute the request only after policy returns allow.[5]
9. Write a detailed audit event regardless of allow or deny outcome.[25][16]
## Discovery and metadata
The system should support a discovery model similar in spirit to OpenID Connect metadata publication, where a server publishes machine-readable configuration at a well-known location.[33][36]
A trust domain or registry should publish metadata that includes:
- issuer identifier,
- supported token formats,
- public key set URL,
- registry URL,
- resolution URL,
- supported protocols,
- supported policy hints,
- approval and revocation endpoints.[33]
This enables automated client configuration and lowers integration complexity.[33][12]
## Compatibility requirements
To support broad adoption, the specification should define compatibility profiles:
- **OIDC profile** for environments with conventional enterprise identity providers.[12][36]
- **SPIFFE profile** for workload-centric infrastructure and service identity.[10][18]
- **Developer profile** for local or low-assurance environments using simpler bootstrap mechanisms during migration.[13][37]
- **Hybrid bridge profile** for organizations mapping existing service accounts, IAM roles, or API keys into delegation flows.[13][3]
## Operational requirements
The specification should support the following operational features:
- revocation endpoint,
- emergency deny list,
- approval callbacks,
- policy simulation mode,
- audit export,
- interoperability tests,
- version negotiation.[26][27][38]
## Open issues
- How attestation should be standardized across providers remains unsettled.[30]
- How cross-registry federation should work is still an open design question.[22][6]
- The minimum common policy language is not yet fixed.[19][20]
- The trust score model should remain optional until its semantics are less ambiguous.[34][39]
## References
[1] Google, "Agent2Agent (A2A) Protocol," https://github.com/google-a2a/A2A, 2025.
[2] Anthropic, "Model Context Protocol (MCP)," https://modelcontextprotocol.io, 2024.
[3] Cloud Native Computing Foundation, "SPIFFE: Secure Production Identity Framework for Everyone," https://spiffe.io, 2021.
[4] IETF RFC 7519, "JSON Web Token (JWT)," May 2015.
[5] NIST SP 800-207, "Zero Trust Architecture," August 2020.
[6] IETF Draft, "OAuth 2.0 Rich Authorization Requests," RFC 9396, October 2023.
[7] IETF RFC 8693, "OAuth 2.0 Token Exchange," January 2020.
[8] M. Sahraei et al., "USENIX Security: Ghostbusters: Defeating the Malicious Agent Problem," 2024.
[9] IETF RFC 7662, "OAuth 2.0 Token Introspection," October 2015.
[10] CNCF SPIFFE, "SPIFFE Workload API," https://github.com/spiffe/spiffe, 2020.
[11] IETF RFC 6749, "The OAuth 2.0 Authorization Framework," October 2012.
[12] OpenID Foundation, "OpenID Connect Core 1.0," November 2014.
[13] NIST SP 800-63B, "Digital Identity Guidelines," June 2017.
[14] NIST SP 800-190, "Application Container Security Guide," September 2017.
[15] IETF RFC 9700, "Best Current Practice for OAuth 2.0 Security," January 2025.
[16] IETF RFC 8705, "OAuth 2.0 Mutual-TLS Client Authentication," February 2020.
[17] Anthropic, "MCP Tool Use Specification," https://modelcontextprotocol.io/docs/concepts/tools, 2024.
[18] CNCF, "SPIFFE/SPIRE Documentation," https://spiffe.io/docs, 2023.
[19] OASIS, "XACML 3.0 Standard," January 2013.
[20] Open Policy Agent, "OPA Documentation," https://www.openpolicyagent.org/docs, 2023.
[21] NIST SP 800-162, "Guide to Attribute-Based Access Control," January 2014.
[22] IETF RFC 7591, "OAuth 2.0 Dynamic Client Registration," July 2015.
[23] IETF RFC 7517, "JSON Web Key (JWK)," May 2015.
[24] IETF RFC 9396, "OAuth 2.0 Rich Authorization Requests," October 2023.
[25] NIST SP 800-92, "Guide to Computer Security Log Management," September 2006.
[26] IETF RFC 7009, "OAuth 2.0 Token Revocation," August 2013.
[27] IETF RFC 8414, "OAuth 2.0 Authorization Server Metadata," June 2018.
[28] GitHub, "Mermaid diagram support in Markdown," https://github.blog/changelog/2022-02-14-include-diagrams-in-your-markdown-files-with-mermaid/, 2022.
[29] C. Nielsen, "Mermaid: Generation of Diagrams and Flowcharts from Text," https://mermaid.js.org, 2023.
[30] TCG, "Trusted Platform Module (TPM) 2.0 Library," March 2014.
[31] OpenAPI Initiative, "OpenAPI Specification 3.1.0," https://spec.openapis.org/oas/v3.1.0, 2021.
[32] JSON Schema, "Understanding JSON Schema," https://json-schema.org/understanding-json-schema, 2023.
[33] IETF RFC 8414, "OAuth 2.0 Authorization Server Metadata," June 2018.
[34] IETF RFC 6819, "OAuth 2.0 Threat Model and Security Considerations," January 2013.
[35] IETF RFC 7523, "JSON Web Token (JWT) Profile for OAuth 2.0 Client Authentication," May 2015.
[36] IETF RFC 8628, "OAuth 2.0 Device Authorization Grant," August 2019.
[37] IETF RFC 7636, "Proof Key for Code Exchange by OAuth Public Clients (PKCE)," September 2015.
[38] IETF RFC 7235, "Hypertext Transfer Protocol (HTTP/1.1): Authentication," June 2014.
[39] NIST, "AI Risk Management Framework (AI RMF 1.0)," January 2023.