Spec Extension: Managed Agent Identity Binding (did:trail:agent:* + PlatformIdentityBinding VC)

[chrishooooo-netizen]

Overview

This issue tracks the v1.2 spec extension for platform-hosted AI agent identity.

Context: Anthropic Managed Agents (launched April 9, 2026) and similar platform-hosted agent runtimes (Azure AI, Google Vertex) challenge a core assumption of the current spec: that an agent has a stable, persistent identity and can directly create its own DID.

In practice, platform agents are dynamically provisioned per session. The persistent entity is the deployment (a configuration), not the running instance.

Proposed Extensions

1. Agent Deployment Identity (did:trail:agent:*)

New identifier mode for agent deployments, registered by the deploying organization:

did:trail:agent:{deployment-suffix}

• Registered by the deployer org (which holds did:trail:org:*)
• Represents one deployment configuration across all instances
• Lifecycle tied to the active deployment, not individual sessions

2. PlatformIdentityBinding VC

New VC type linking a platform's internal deployment ID to a did:trail:agent DID, signed by the deployer — not the platform.

Key design: no platform cooperation required for external audit. A regulator verifying an EU AI Act Art. 12 audit trail does not need to contact Anthropic, Azure, or Google.

Open Questions

1. Should the PlatformIdentityBinding be self-attested (deployer only) or require third-party countersignature for Annex III high-risk deployments?
2. Mid-session revocation for long-running agents (6+ hours) — push notification to MCP servers, short-TTL attestation, or out-of-band?
3. Cross-registry scenarios: Anthropic agent calling Azure service, both registered in did:trail — how does cross-resolution work?

Status

Spec draft in progress. Community input welcome on the open questions above before normative text is written.