Execution profiles
VACT-P supports three execution profiles over the same trust model: interactive synchronous sessions, durable asynchronous work, and ambient agents operating under renewable background mandates. A fourth cross-cutting protocol governs any irreversible or high-risk action.
Synchronous profile
For interactive assistants, real-time collaboration, coordinated robots, and low-latency agents. A session must carry a session ID, participants, negotiated protocol version, authenticated channel, expiry, sequence state, cancellation mechanism, and budget state.
Streaming frames
A DELTA is never a final contracted output unless the contract explicitly permits partial settlement.
Coordination primitives
BARRIER
Wait for a named participant set.
QUORUM
Proceed after a threshold signs an identical proposal digest.
LOCK
Reserve a resource for a bounded lease.
EPOCH
Identify a coordination round.
ABORT
Cancel the current round.
COMMIT
Authorize external effect.
Locks expire automatically and should use fencing tokens to prevent stale holders from acting.
Asynchronous profile
For durable work that may outlive a connection or process.
Processing lease
A worker must acquire a bounded lease before executing a queued task: task ID, worker runtime identity, lease epoch, issue time, expiry, allowed checkpoint, and proof. Only the holder of the highest valid lease epoch may commit progress.
Checkpoint
A checkpoint should include the completed step, an output commitment, consumed budget, a remaining-plan summary, a resumability data reference, and runtime/model identifiers — all subject to the contract's privacy and retention policy.
Retry
Retry policy specifies maximum attempts, backoff, retryable reason codes, a deadline, whether a new worker is permitted, and whether prior partial artifacts may be reused. Economic settlement must deduplicate repeated attempts and charge only according to the Task Contract.
Ambient profile
Ambient agents observe events and act under a renewable, observable authority lease — never a silent, indefinite one.
Ambient Mandate
Must specify: subscribed event types, permitted and prohibited actions, rate limits, notification limits, quiet periods, monetary and compute budgets, data policy, renewal interval, absolute maximum lifetime, escalation conditions, approval rules, a kill-switch endpoint, and a visible activity-feed endpoint.
An ambient agent must never subscribe to new event classes without authority, infer permission from prior silence, expand its own budget, alter its own renewal policy, suppress its activity feed, or delegate ambient authority unless explicitly allowed.
High-risk action protocol
An irreversible or high-risk action uses a staged protocol whenever policy requires it — proposal, preparation, approval, and commit are separated so an approved action can be bound to an exact digest before anything happens.
Acting Agent → Policy Engine : PROPOSE(action digest)
Policy Engine → Acting Agent : PREPARED(policy decision, constraints)
Acting Agent → Approver : APPROVAL_REQUEST
Approver → Acting Agent : APPROVAL_SIGNATURE
Acting Agent → Resource : COMMIT(action, approval, idempotency key)
Resource → Acting Agent : EFFECT_RECEIPT
Acting Agent → Approver : EXECUTION_RECEIPT| Phase | Purpose |
|---|---|
PROPOSE | Describes the intended effect without performing it. |
PREPARE | Validates authority, policy, budget, and resource readiness. |
APPROVE | Collects required human or organizational approval. |
COMMIT | Performs the authorized effect exactly once where possible. |
RECEIPT | Records the observed result. |
A PREPARE authorization expires after a short interval and is bound to the exact action digest — it cannot be reused for a different or modified action.