Authority that can be traced and bounded.

Represent where permission to cause an outcome came from, how it was delegated and why it remains valid.

Control plane / conceptual
  1. Attributable principal
  2. Valid authority origin
  3. Current policy and trust
  4. Enforceable action decision
Conceptual authority chain
  1. 01
    OriginOrganization

    Accountable task authority · $100K ceiling

  2. 02
    DelegationFinance agent

    Specific workflow · $25K delegated

  3. 03
    Child grantInvoice agent

    Validate and prepare only · expires with task

  4. 04
    Requested effectPayment tool

    release_payment · $8,400

Action-time decisionDENY

Payment release was not delegated to the acting agent.

Parent authorityApproved invoices ≤ $25,000
Child taskMatch documents for invoice I-42

Payment release remains outside the child grant. The authority chain above shows the action-time DENY.

Authority origin

Authority has an origin

An autonomous action should be traceable to an organization or person with the right to authorize it. The origin is more than a credential issuer: it is the source of the permission that can be delegated. A grant identifies a principal, permitted purpose, scope, constraints and expiry. It may also state whether sub-delegation is allowed.

Without an origin, a valid agent identity can become a blank check. A service account may authenticate successfully while its requested action has no connection to the task a human approved. The authority model separates those facts.

  • Originating organization or human authorization
  • Named principal receiving the grant
  • Purpose and permitted action scope
  • Constraints, expiry and revocation state
  • Explicit delegation rights

Delegation

Delegation forms a graph

A parent agent may assign a narrower task to a child agent. That child grant must be explainable as a subset of the parent’s authority. If the parent may validate invoices but not release payments, the child cannot acquire payment authority simply because it has a payment-tool credential. The authority graph records parent and child relationships so the decision can be traced back to its origin.

Monotonic narrowing is the governing rule: each delegation can retain or reduce scope, shorten duration, add conditions or remove sub-delegation rights. It cannot widen the parent grant. This makes a chain of agents reviewable even when tasks fan out across runtimes.

Purpose, scope and expiration

Authority should describe what the task is for, which resources or actions it covers and when it ends. Scope might include one account, API operation, machine, geographic area, counterparty or amount. Purpose distinguishes a legitimate workflow from the same API call made for an unrelated task. Expiry avoids indefinite privileges after a job finishes.

These dimensions are evaluated together. A grant to adjust a cooling set point may be valid for a particular data center during a maintenance window, but not for a neighboring facility or after the window closes. A parent’s revocation should invalidate dependent child authority, even if their own timestamps have not expired.

Revocation

Revocation and explainable denial

Authority changes after issuance. A person can cancel a task, a principal can be suspended, a parent grant can be revoked, or an operating state can make the action impermissible. A runtime decision should use the current graph rather than assume an old grant is still valid.

A denial should identify the failed boundary: no origin, expired grant, scope mismatch, forbidden sub-delegation, revoked parent or policy conflict. That explanation lets operators fix authorization design without giving agents more standing privilege than they need. ProofGrid traces authority through the action decision. The exact interface and enforcement topology must be confirmed for each deployment.

What a grant should be able to answer

A useful authority grant is not merely a role label. It should answer which principal received authority, which origin issued it, what operation it covers, what resource or target is in scope, why the task exists, when the authority expires and whether the principal may delegate further. Constraints may include value, geography, counterparty, operating mode and a required approval. The exact representation can vary, but each fact must be available to the decision that governs a consequential action.

Consider a procurement task. “Purchasing agent” is an identity or role; it says little about the current request. A grant that permits drafting purchase orders for approved suppliers in a named business unit until a deadline is more precise. The authority to submit a purchase order may be separate. The authority to pay it may belong to a different principal entirely. This separation protects against a common failure in agent systems: a task-specific instruction becomes a general privilege because the tool credential is broader than the human authorization.

Authority across multiple systems

A workflow can cross an agent runtime, an API gateway, a queue and an execution service. The grant must remain attributable through those handoffs. A downstream service receiving only a shared workload token may be unable to tell which agent and task originated the request. The architecture therefore needs a way to carry or resolve the relevant authority context at the last enforceable boundary, while avoiding untrusted self-assertions from the agent.

A parent-child graph also needs to handle cancellation and concurrency. If a parent has authority for a total of ten changes, spawning ten children must not accidentally give each child ten independent changes. If the parent task is withdrawn, children should not keep acting simply because their local credentials have not expired. These are implementation questions to resolve in an integration design; they are not solved by adding a parent ID to a log after execution.

Authentication, authorization and autonomous authority

Authentication establishes which person, workload, machine or logical agent is presenting a request. Authorization checks whether that principal may access a resource or operation. Autonomous authority asks a further provenance question: who authorized this actor to cause this particular outcome, for this task, under these limits? All three checks are needed. An authority claim from an agent is not proof of authority; the origin and every delegation must be established by trusted systems or accountable approvers.

OAuth scopes are useful for bounding token access to an API. RBAC organizes stable permissions by role. ABAC evaluates attributes of subject, resource and environment. Capability-based systems can convey constrained rights. None should be dismissed: each can implement part of a sound design. A long-running agent workflow adds task purpose, delegation lineage, action parameters, revocation and shared limits that a broad token or role alone may not express. A capable architecture layers these controls rather than replacing them.

Conceptual Authority Grant

A conceptual grant may record principal, accountable origin, parent-grant reference, task purpose, permitted action, resource or target, constraints, valid-from time, expiry, delegation permission and depth, approval requirement and revocation state. This is a design model, not a finalized ProofGrid public API schema. Its purpose is to expose which facts an action-time decision must verify.

An origin should be issued or approved through an accountable channel. A child grant must refer to a verifiable parent and carry no wider action, target, value, time or delegation right. At each cross-system handoff, the receiving component must bind the request to trusted identity and grant context. Accepting a free-form parent ID or an agent-written explanation as authorization invites forged authority.

  • Principal and accountable origin
  • Parent reference and delegation depth
  • Purpose, action and target
  • Constraints and shared limits
  • Valid-from, expiry and revocation
  • Approval requirement and provenance

Integrity of an authority chain

Monotonic narrowing is the core delegation invariant: a child can receive a subset of the parent, never an expansion. Sub-delegation requires its own permission and a depth bound. If two upstream approvals are required, effective authority is their intersection, not their union. When five children share a parent budget of $25,000, each cannot independently spend $25,000. Possible designs include one atomically updated remaining budget, reservations against it or explicit partitions. The choice depends on concurrency and failure semantics.

Expiration should be short enough for the task and checked at use, not only at grant creation. Revoking a parent must invalidate dependent descendants at their next consequential request. An approval or authorization artifact for a high-impact action should bind actor, operation, target and normalized parameters; it should be short-lived and resistant to replay. A new target or changed amount requires a new decision. The architecture must also define what happens when the chain cannot be verified: high-consequence writes normally stop, while a separately scoped read-only mode may remain available.

Platform / Next Step

Make authority explicit at the action point.

Review an Authority Model