RU
Reference/Foundations

Autonomy and accountability

Define delegated authority, trace decisions, and establish how uncertain outcomes are resolved.

Proposal2 min readEdition 0.1
MarkdownJSON

#Authority is specific

Autonomy should be described per activity. An organization might select research questions independently while requiring approval before external publication. It might schedule work freely but spend only within a fixed allocation. These boundaries reveal substantially more than a single label such as supervised or fully autonomous.

Accountability begins with identifying the party responsible for granting authority and the system responsible for exercising it. A mandate records intended limits; technical controls enforce available limits. Neither a confident model answer nor a successful tool invocation establishes that an action was authorized or that it achieved its purpose.

#Trace decisions and effects

A decision record should connect an instruction, the applicable authority, the action attempted, the observed result, and the basis for acceptance. The appropriate detail depends on the action. Correcting a private draft and making a public commitment require different levels of evidence and review.

Some effects remain uncertain after a timeout or crash. The organization should preserve that uncertainty and reconcile external state before repeating consequential work. A durable record of intention helps recovery, but it cannot establish whether an external service completed an operation whose response was lost.

#Escalation and correction

Escalation is an operating mechanism for decisions outside authority, ambiguous ownership, missing resources, or unresolved evidence. A useful escalation identifies the blocked decision, the facts already established, and the precise choice required. It should leave enough state for authorized work to continue where appropriate.

Correction also needs an owner and a scope. The ability to revise procedures is different from the right to expand permissions. Independent review may strengthen assurance, but its independence must be described: another agent using the same model and evidence can still share the original error.

Sources & context

  1. Wooldridge, Jennings & Kinny (2000). The Gaia Methodology for Agent-Oriented Analysis and Design
  2. Hübner, Sichman & Boissier (2002). A Model for the Structural, Functional, and Deontic Specification of Organizations in Multiagent Systems
  3. Fourney et al. (2024). Magentic-One

Reference edition 0.1 · 2026-10-14. Read the editorial method.

Continue exploring

Defining properties

A proposed set of organizational properties and the evidence needed to recognize them.

Evaluation methods

Measure organizational outcomes, continuity, intervention, and resource use with explicit scope.

Editorial method

How Axiom AO distinguishes evidence, proposed definitions, examples, and changing knowledge in two languages.