Skip to content
Engineering Insight
An Approval Prompt Is Only One Part of Delegated Authority
Research

An Approval Prompt Is Only One Part of Delegated Authority

Pinned source-function probes show how approval changes action disposition, while leaving caller identity, authority scope and policy coverage as separate evidentiary questions.

2026-11-12中文 →

An Approval Prompt Is Only One Part of Delegated Authority ​

After a human approves, the executor continues the original action. That is a specific promise that code can test. Saying that a delegated invocation was correctly authorized requires more: who initiated it, what the approval covered, and whether the action passed the applicable checks. A confirmation prompt cannot supply all those connections by itself.

This study places that distinction against a bounded execution result. In selected functions from a pinned AgentSpec revision, agreement continues the same action, refusal skips it, and empty or nontriggering rules preserve the action. Our interpretation is that action approval, policy coverage and caller authority need separate evidence. This neither certifies a complete platform nor alleges a bypass in a deployed product.[1][2]

What actually continues after confirmation ​

In the inspected synchronous source path, constraint checking precedes actual action execution. The human-inspection function receives an action, asks the user and returns a disposition. The same-day analysis extracted and executed selected functions at commit e6fa3902e2cfb9681f454b355691b771f70543f8. This is the registered baseline revision, not an assertion about current HEAD or the exact paper release.[1]

The internal experimental evidence covers five approval-input cases and six dispatch cases. Action, Rule, RuleInterpreter and state were disclosed test doubles. Real domain predicates, adapters and LangChain were not executed. This is a source-derived isolated function probe, not a full-system reproduction.[2]

Input or conditionInternal probe observationSupported reading
yes, or y surrounded by whitespaceContinue with the same action objectAgreement does not replace the action in this function
no or nSkip with a skip actionThe function can suppress this action
Invalid input followed by yesTwo prompts, then the original actionInvalid input does not directly authorize continuation
Triggered rule with injected continue, skip or stop resultOriginal action, skip action or finish action respectivelyDispatch follows the supplied interpreter result
Nontriggering rule or empty rule listOriginal action retainedSuppression depends on coverage and triggering
Missing rules valueException stating that rules must not be noneAn empty list and absent rules are different paths

All table entries are internal isolated-function observations, not independent third-party validation or samples of real user invocations. Continue, skip and stop outcomes were supplied by the interpreter double. Correct dispatch therefore says nothing by itself about whether a real predicate evaluates correctly.[2]

Object identity establishes only a return relationship in this synchronous call. The probe did not test concurrent mutation, changes after approval, asynchronous calls or recovery. Translating “the same object continues” into “approved content always equals executed content” would exceed the evidence.

A check exists; coverage remains a separate question ​

Empty and nontriggering rules are necessary controls. They show why the presence of a checking call does not establish that every action requiring restriction is caught by an effective rule. Assessing protection for an action still requires the applicable rule, its trigger, the state read by its predicate and the eventual execution point. Only a small part of that chain was executed here.

The AgentSpec paper describes trigger, predicate and enforcement mechanisms, with checks at action, state and finish boundaries; the reported integration uses LangChain 0.3.13. That description supports discussion of where to inspect. It cannot supply missing audits of interpreters, domain adapters or asynchronous paths.[3]

The paper also reports costs and differences in coverage: safe embodied-task fulfillment falls from 58.62% to 54.26%; generated rules prevent five of eight driving violations, while manual rules achieve complete compliance in the tested scenarios. These are author-reported results under the stated experiments, not our reruns. They motivate evaluating rule selection rather than reading the existence of a constraint mechanism as a guarantee that every task completes safely.[3]

Organizational relationships, tokens and answers do different jobs ​

Salesforce's Help article describes parent-planner delegation within one organization, explicit relationships across organizations, and per-agent gateway controls. These are vendor capability descriptions. The captured material does not provide the delegated credential scope, token trace or execution evidence needed to verify an individual invocation. Organizational relationships alone cannot establish that each tool call has been correctly authorized.[4]

The captured MCP draft addresses a different responsibility. Authorization is optional in that draft; HTTP implementations supporting it should follow the specified flow, while STDIO should obtain credentials from the environment. For requests using the authorization flow, resource servers must validate tokens for their own audience. The security requirements additionally call for separate upstream API tokens and prohibit passing the client token through.[5][6]

The captured security guidance treats state handles as ordinary tool arguments. They cannot authenticate callers and should be bound server-side to a verified principal. Guessing a handle should not authorize access to another user's state. This is a normative obligation, not a measurement that a particular service implements it correctly.[7]

These materials can be compared by responsibility, but cannot be assembled into an integration test that never happened. Organizational relationships describe collaboration; token audience and upstream credential restrictions define resource boundaries; state binding concerns ownership of state; a human answer decides the action at the observed checkpoint. Only the last received isolated branch execution here. Descriptions and obligations in the other layers cannot substitute for that execution evidence, and the local probe cannot validate those layers in return.

Our engineering interpretation is to ask separately how the caller is established, how action and resource scope are restricted, which action occurrence an approval refers to, and whether all actual execution paths encounter the applicable rules. These are inspection questions derived from the evidence, rather than an implemented or validated complete authorization design.

Evidence and sources ​

Public primary-source facts here describe what a specification or pinned source says. Source-reported claims remain attributed to vendors and authors. Our observations and internal experiments cover only inspected code and selected executed functions. Our interpretation separates responsibilities that need verification. No independent third party examined the complete delegation chain discussed here. Citation is not endorsement, peer review is not proof, and self-reported evidence is not independent evidence. Publication or successful implementation does not establish general validity.

  1. AgentSpec authors, pinned source files controlled_agent_excector.py and enforcement.py, registered baseline commit e6fa3902e2cfb9681f454b355691b771f70543f8, local snapshot read on 2026-11-12. Executor; human inspection.
  2. Research Center, internal isolated-function experimental record, 2026-11-12: five approval and six dispatch cases. Disclosed doubles, inputs, outputs and limits are retained in the experimental record bound to the same-day Research Object. This is neither a public-product test nor third-party reproduction.
  3. AgentSpec authors, Customizable Runtime Enforcement for Safe and Reliable LLM Agents, research paper v3, mechanism and Tables 4–6, snapshot read on 2026-11-12. Paper.
  4. Salesforce, Help documentation on single-organization and multi-organization collaboration and Agent Gateway, snapshot read on 2026-11-12. Help article.
  5. MCP project, draft authorization specification, scope, protocol requirements and token handling, captured on 2026-11-12; a mutable draft. Authorization.
  6. MCP project, draft authorization security considerations, audience binding and privilege restriction, captured on 2026-11-12. Security requirements.
  7. MCP project, draft security best practices, state-handle hijacking guidance, captured on 2026-11-12. Security guidance.

A local check may suffice; distributed bindings remain unestablished ​

A synchronous local approval can be entirely adequate for a single-user threat model. Absence of distributed approval records in the inspected files does not establish that the complete system lacks those controls. Untested paths cannot be described as vulnerabilities.

What remains unestablished is how a child action binds to its initiating principal, task occurrence and approval occurrence; how revoked grants affect long-running delegation; and whether asynchronous, tool-internal and recovery paths encounter equivalent constraints. Answering those questions requires evidence from the relevant paths and credentials. The present boundary is narrower: approval changed action disposition in the tested functions, while delegated authority still needs its own supporting evidence.

Last updated: