Anthropic defines the agent.
VAIG decides what it may do.
What Anthropic published
An agent identity and access model for Claude. Agents receive workspace-level identities rather than impersonating individual users, with scoped per-channel credentials, restricted outbound traffic, and audit logs of actions taken.
Future roadmap item: just-in-time credential grants.
Two different questions
Anthropic — Identity
Layer 1–2 work: authentication, credential management, access scoping.
"Who is this agent, and what systems may it touch?"
VAIG — Execution Control
Layer 3 work: pre-execution admissibility gating, authority delegation, WORM proof.
"What may it actually do right now — and can you prove it happened?"
Where they diverge
| Question | Anthropic (Identity) | VAIG (Execution Control) |
|---|---|---|
| Who is the agent? | ✓ Workspace identity | — |
| What systems may it access? | ✓ Channel-scoped credentials | — |
| Is this action authorized now? | ✗ Not addressed | ✓ AARM 6-primitive gate |
| Under whose authority was it delegated? | ✗ Not addressed | ✓ RRP + authority chain |
| Can you prove it happened? | Audit log (append-only) | ✓ SHA-256 WORM receipt (hash-chained) |
| Escalation mid-task? | ✗ Future roadmap (JIT grants) | ✓ STEP_UP primitive — live today |
Key insight: Anthropic's JIT credential grants are listed as a future feature. VAIG's STEP_UP primitive — live escalation to stronger authority on demand — already implements this. The execution control layer is the recognized next problem. VAIG is 18 months ahead.
"Anthropic tells the agent who it is.
VAIG decides what it may do — and proves it happened."
Competitive stack — updated
Anthropic now spans L1 (agent identity) and L4 (model). They are building toward the middle but have not crossed into L3. VAIG's deterministic pre-execution gating + formal verification + WORM receipt remains architecturally distinct.
Strategic implications
- Anthropic's identity model is infrastructure VAIG builds on top of, not competition
- Combined stack: Anthropic identity + VAIG execution control = complete governed agent architecture
- Pitch to enterprise: "Your agent already knows who it is. Now you need to control what it does."
- Pitch to Anthropic ecosystem partners: VAIG is the execution control complement to Claude's identity layer
- Layer 3 execution control remains uncontested — 18-month window holds