Durable Agent Authority — for agents that sleep, retry and resume

Durable-execution frameworks make an agent's work survive sleeps, retries, crashes and restarts. KYE Protocol™ makes sure its authority does not silently survive with it. A workflow resumed after thirty days may resume with an expired delegation, a revoked scope, a changed policy version, or a human approval that covered the start of the run but not its continuation. Durability of execution is not durability of authority.

KYE™ is not a durable-execution framework, a workflow engine or an orchestrator. It is the authority layer underneath them: at every resume, retry and continuation, KYE™ re-resolves whether the agent still holds valid authority for its next consequential action, then emits a signed, replayable record. Your durable runtime asks whether the workflow survived; KYE™ asks whether its authority did.

Start a durable-agent authority pilot See the governed agents

The line: a resumed workflow vs. a re-authorised action

Your durable runtime replays event history and restores state so the agent continues exactly where it stopped. That is the correct behaviour for the work — and the wrong default for the authority behind it. Authority has an expiry, a scope, a policy version and an approver, and all four can change while the workflow sleeps.

Your durable runtime asks Did the workflow survive the crash, the sleep, the retry — and can it continue deterministically from recorded state?

KYE™ asks When it continues, does the agent still hold valid authority — unexpired mandate, unrevoked scope, current policy version, fresh-enough approval — for the next consequential action?

A durable agent that resumes is a new authority event, not a continued one. The run identity persists; the permission to act must be earned again at the boundary. KYE™ answers with allow / deny / require_approval / suspend, and every verdict seals a KYE Evidence Pack™ that an auditor can replay long after the run has finished.

Every one of these is a new authority moment

Long-running agents cross authority boundaries that short-lived ones never see. KYE™ gates each of them at the moment it occurs — not once at launch.

Resume after sleep — the delegation that authorised step one may have expired before step nine wakes up.
Retry after failure — a retried payment or filing is a second consequential action, not a replay of the first.
Policy version drift — the rules changed mid-run; the next action must resolve against the current version, with the change on the record.
Approval staleness — a human sign-off at launch does not cover a continuation weeks later; KYE™ re-checks whether the approval still binds.
Revocation mid-run — when authority is withdrawn, the running workflow must lose its permission at the next boundary, not at the next deploy.
Handover between agents — a continuation picked up by a different worker or sub-agent is a new principal and needs its own delegation chain.

This is continuous authority resolution: the mandate is checked at every consequential step for the life of the run, and each verdict is signed, tenant-scoped and replayable.

For teams building on durable-execution frameworks

If your agents run on Temporal-style durable workflows — resumable state machines, event-sourced steps, human-in-the-loop pauses — KYE™ is the authority layer beneath the framework. The framework owns progress; KYE™ owns permission. This division is deliberate: KYE™ never runs, schedules or orchestrates the workflow (§0.30 boundary) — it decides whether each step's consequence may become final.

The integration is a runtime check at the step boundary: before a consequential step executes — on first run or on resume — the agent asks KYE™ for a verdict. KYE™ evaluates actor, delegation chain, purpose, data scope, policy version, approval freshness and finality state, returns the decision, and seals the evidence. Your workflow code stays deterministic; the authority state lives outside it, where it belongs — an approval baked into workflow history at launch cannot express an authority that changed while the run slept.

One authority spine under every durable agent

Durable-agent governance reuses the same runtime primitives that govern every other KYE™ action — one control point, whether the agent lives for a second or a quarter. Nothing here is a separate durable-agent system.

Authority Gateway™Runtime allow / deny / require approval / suspend at every step boundary, including resume.

Purpose Permission™Binds each step to a permitted purpose and data scope — re-checked on continuation.

Decision Map™Inputs, policy version, model, human override and verdict for each step's decision.

Evidence Pack™Signed proof of actor, mandate, approval freshness and finality — one per verdict, for the life of the run.

Replay-Proof™Offline reconstruction of why any step was allowed or blocked, from public keys alone — months after the run ended.

Entity & Principal RegistryEvery durable worker, sub-agent and continuation is a governed principal with its own delegation chain.

The result is one provable answer for every step of a long-lived run: who held authority at that moment, for what purpose, under which policy version, with what evidence — and whether the step became final under Authority Finality™.

Start a durable-agent authority pilot Browse the governed agents