This note proposes AgentIDL, an implementation-independent interface layer for agent systems.
As AI systems are increasingly deployed in production and industrial contexts requiring stable behavior across heterogeneous environments, AgentIDL identifies the need for a stable semantic boundary to support constraint, audit, and interoperability at the level of intent.
This document is a draft produced by a W3C Community Group. It is not a W3C Recommendation and does not represent consensus of the W3C Membership.
As agent systems proliferate, current implementations commonly collapse natural language interpretation, intent representation, execution structure, and runtime control into a single layer.
This collapse impedes interoperability, auditability, and responsibility attribution. The absence of an explicit interface layer between intent and execution constitutes a structural gap in current agent architectures.
AgentIDL is proposed to address this gap.
For the purpose of this Community Group:
An agent is an entity whose actions are subject to constraint and audit.
This definition is intentionally minimal and does not assume autonomy, intelligence, interaction model, or internal reasoning structure.
An entity that cannot, even in principle, be constrained or audited does not fall within the scope of this work.
AgentIDL defines a semantic interface boundary between what is intended and how it is executed.
It does not prescribe reasoning strategies, execution mechanisms, or control-flow models.
AgentIDL functions as a semantic contract that enables intent to be expressed, constrained, and examined independently of implementation.
AgentIDL is aligned with WebIDL at the level of architectural intent.
WebIDL establishes stable interfaces between web-exposed semantics and heterogeneous browser engines without being defined by any single implementation.
AgentIDL applies the same principle to agent systems: interface-first design, implementation independence, and interoperability grounded in shared meaning.
AgentIDL is not an agent framework and is not an agent protocol.
Frameworks and protocols operate at the level of implementation and interaction. AgentIDL operates at the level of intent-to-execution interface definition.
Compatibility with existing systems may exist, but such compatibility is derivative rather than definitional.
Current agent systems typically transition directly from natural language or internal reasoning to framework-specific execution logic.
This transition lacks an explicit, stable interface. AgentIDL introduces this missing layer by making intent representation, constraints, and auditability explicit and independent of runtime.
At present, no mainstream agent implementation explicitly operates at this layer.
The primary interoperability concern addressed by AgentIDL is natural language intent alignment.
Interoperability is evaluated by preservation of intent boundaries, stability of meaning across execution contexts, and resistance to semantic drift introduced by implementation assumptions.
Proposals to adjust the scope of AgentIDL must be justified at the level of socio-technical compatibility.
Arguments based solely on implementation convenience or framework design do not constitute sufficient grounds for scope modification.
Maintaining implementation neutrality is a prerequisite for interoperability.
AgentIDL formalizes an architectural boundary that is currently implicit and unstable.
By defining a semantic interface between intent and execution, grounded in constrainability and auditability, AgentIDL provides a foundation for interoperable and governable agent systems.
This note invites discussion on the role and shape of this missing layer, without constraining future implementations.