It acts at machine speed inside an envelope you wrote. High-risk actions wait for a person. Nothing goes unlogged.
{
"decision": "revoke_idle_token",
"actor": "agent",
"policy": "nhi-envelope@v3 / machine_speed",
"subject": "svc-vendor-sync",
"reason": "no activity for 41 days (limit 30)",
"evidence": "sha256:9f3a…c41e",
"prev": "sha256:71c0…08bd",
"reviewer": null
}
Sample data. A machine-speed decision has no reviewer; a checkpoint decision names one.
Autonomy is only trustworthy when its boundary is explicit. These limits are written into the policy model, and the agent runs inside them.
Every action class is assigned a mode before the agent acts on it: machine-speed, checkpoint or human-gated. The agent cannot reclassify an action, widen its own scope or grant itself access. Changing the envelope is itself a human-gated, logged event.
Every decision, automatic or human, including the actions it chose not to take, is written to the ledger with a chain of custody as it takes effect.
Vault items are encrypted on the device with AES-256 and stay ciphertext in Sequrely’s cloud. Decisions are made on signals and policy, not on your secrets.
Deletions, vault exports, scope widening and access to systems you flag as critical are human-gated by default. The agent cannot move them out of that class.
The digital twin learns a baseline in observe mode first. Enforcement starts when you turn it on, per action class, not when the agent decides it is ready.
Three modes, assigned per action class. An example starting envelope.
Reversible, low-blast-radius actions: ending idle sessions, rotating an expired share, containing an identity whose drift score crosses the line you set. Executed in seconds, sealed, reviewable.
The agent prepares the decision and the evidence, then waits. A named approver signs off in one tap before it runs. Just-in-time grants and new-device authorizations live here.
Deletions, exports, scope changes, policy changes and access to systems you flag as critical. The agent can recommend; only a human can execute. Nothing moves this class down.
Agents get twins too; theirs are tighter, because a deploy bot has no reason to be curious.
A deploy bot’s token is presented from a network it has never used.
The twin flags token reuse plus new context. Risk crosses the envelope’s threshold.
revoke_on_token_reuse is a machine-speed action in this scope. No human is needed yet.
The token is revoked and the credential rotated. The CI runner picks up the new secret.
Inputs, score, mode, action and actor are written and hash-chained. The security lead sees it in the ledger, not in a 03:00 page.
Agent and human rows share one ledger. That is what makes bounded autonomy inspectable.
Sample data: investigation ledger. Example row: 09:41:02, Sequrely agent, Ended idle session · svc-payroll-bot · idle 31m, status Revoked.
Sample data. Identifiers are fictional; humans are shown as roles.
Targets the enforcement engine is built to meet.
Targets the enforcement engine is built to meet, measured on sample estates.
Low-risk decisions execute at machine speed. High-risk actions always route to a human.