Status: Draft Specification: Agent Contracts v1.1
This document defines the terminology used by the Agent Contracts specification.
The purpose of this document is to ensure that core terms have consistent meanings across:
- contract authors,
- validators,
- governance systems,
- policy engines,
- runtime integrations,
- documentation,
- and independent implementations.
A term defined here SHOULD be interpreted according to this document unless a specification version explicitly defines a more specific meaning.
Terminology is part of the semantic foundation of Agent Contracts. Changes to the meaning of a core term SHOULD therefore be treated as specification-level changes rather than ordinary documentation edits.
An Autonomous System is a software or computational system that can perform one or more actions on behalf of a user, organization, or another system with some degree of independent operation.
An autonomous system may:
- receive information,
- make decisions,
- invoke tools,
- access resources,
- modify state,
- communicate with external systems,
- or produce external side effects.
An autonomous system does not need to use a particular AI model, agent framework, workflow engine, or implementation architecture.
Examples may include:
- an AI research assistant,
- an educational tutoring system,
- a financial analysis agent,
- a customer-support agent,
- a software-development agent,
- a business automation system,
- a data-analysis system,
- or a multi-step system containing autonomous decision-making.
The term is intentionally broader than any particular definition of "AI agent."
An Agent is an autonomous system that performs actions or makes decisions on behalf of a user, organization, or another system within a defined operational context.
An Agent may use:
- language models,
- machine-learning models,
- deterministic logic,
- external tools,
- workflows,
- APIs,
- databases,
- browsers,
- physical devices,
- or combinations of these.
The term Agent is used by Agent Contracts as a practical shorthand for an autonomous system that is subject to a contract.
Agent Contracts does not require an Agent to use a particular architecture, model, framework, or implementation technique.
An Agent Contract is a machine-readable declarative specification describing the declared operational characteristics and boundaries of an Agent or autonomous system.
A Contract may describe properties such as:
- capabilities,
- resources,
- actions,
- permissions,
- constraints,
- side effects,
- approvals,
- dependencies,
- state,
- recovery behavior,
- replay characteristics,
- and observability.
A Contract describes the operational boundary of a system.
It does not define:
- the system's internal reasoning,
- its implementation code,
- its private prompts,
- its model weights,
- its execution algorithm,
- or its internal architecture unless explicitly represented as an applicable extension.
A Contract is therefore a specification of operational properties rather than an implementation of the Agent itself.
A Contract Version identifies the version of the Agent Contracts specification against which a contract is intended to be interpreted and validated.
A Contract Version MUST be explicitly identifiable.
For example:
version: 1