Search assets, operations, jobs, agents, skills, and approvals across every connected venue, or run an action.
An agent reaches the edge of its authority and stops. A human decides in the real Inbox, and signs a capability with a real expiry.
A drift monitor watches a cohort metric. When it breaches its threshold the monitor does not change policy and does not ask another agent — it escalates to a person, and its work stops until that person decides. What they grant is a capability with a real expiry, signed with their own key.
Select a venue to run this demo.
Registers two cohort windows, a window pointer, a copy of the venue's HITL skill and one monitor agent — under your identity. Re-running is a no-op; nothing is duplicated.
cohort windows and the reviewed limit live under here
the venue needs this provider's API key in Secrets
empty = provider default
Sign in to run setup.
The agents call the model through v/ops/langchain/openai; the venue needs that provider's API key — manage Secrets.
1 · Drift becomes a governed event
The cohort window moves to week two, where device-reuse velocity has tripled. rk-monitor reads both windows under its own capped authority, reports the numbers, and an ask is raised proposing a temporary raise of the reviewed limit. Its job parks — nothing moves until a person decides.
Watch: The job holds at INPUT_REQUIRED. The wait itself is a recorded state.
Select a venue first.
2 · A human decides, and authority is granted
You answer in the real Inbox as the risk officer and sign the capability with your own device key. An agent cannot answer this — the runtime refuses on identity, not on scope, because resolving an ask confers authority. The parked job then resumes on its own.
Watch: The grant is verified here by the venue's own ucan:verify: your DID as issuer, one narrow capability, a real expiry that lapses by itself.
Looking for how the runtime refuses an agent outright, rather than pausing for a human? That is the Adaptive Risk demo — a policy gate that denies before execution.