Skip to main content
Version: 2026.08.10

Per-Agent Identity

Status: rolling out now across the platform. Identity resolution with staged enforcement is live on the platform Brain (2026-08-10); the strict enforcement rungs activate progressively behind observation windows. See the changelog for the rollout announcement.

Page updated 2026-08-10 against the live platform.

Know which agent did it — cryptographically, every turn. Each governed action carries a signed, per-turn identity assertion, so the operating record attributes work to a specific agent, lab, and session — not just "an AI."

Why per-turn, not per-agent tokens

A long-lived per-agent token is replayable for its whole life, and on a shared machine one process can read a sibling's token and impersonate it. A per-turn signed assertion is different:

  • Non-replayable — each turn mints a fresh assertion; a leaked copy is stale by the next turn.
  • Delegation-shaped — the assertion records who delegated to whom (operator → agent), giving an auditable delegation record rather than an opaque label.
  • Separately-keyed per surface — verifiers for different surfaces refuse each other's assertions, closing cross-surface confusion.

How enforcement rolls out

StageBehavior
ShadowIdentity is resolved and logged on every governed call; zero decision change.
Staged enforce (live now)A resolved identity becomes the authoritative principal. Unresolved local callers are permitted and logged (the operator can never be locked out of their own platform); unresolved hosted callers are denied only when the hosted-strict rung is switched on.
StrictHosted strict-deny and local strict activate after clean observation windows and explicit ratification.

The fail-open-by-design asymmetry is deliberate: local resolution failures never lock the operator out; hosted/tenant resolution failures fail closed, because there isolation is the point.

What you see as a customer

Governed actions in your workspace's operating record carry the acting agent's identity (agent, lab, session, surface) — so review, replay, and audit answer "which agent did this," not just "an agent did this."