What is FIPA-ACL (Agent Communication Language)?

FIPA-ACL (Agent Communication Language) is the standard language autonomous software agents use to exchange messages, published by the Foundation for Intelligent Physical Agents, an association founded in Geneva in 1996 that became the eleventh standards committee of the IEEE Computer Society on 8 June 2005. Its governing idea comes from speech act theory: a message does not merely carry data, it declares the sender's intent. One mandatory field, the performative, separates a request from an assertion, from a proposal, from a refusal, and the recipient reads that declaration before reading the content at all. Around this core FIPA standardized the message structure, a library of communicative acts and eleven interaction protocols, among them the Contract Net and, never past experimental status, two auction formats. It is the most complete precedent today's agent protocols have, and reading it explains why several problems that look new are not.

See also: A2A Protocol (Agent2Agent) · Open standard letting AI agents from different vendors discover each other, delegate tasks and coordinate across distinct enterprise systems.

What a message looks like

The ACL Message Structure Specification (SC00061G, Standard status since December 2002) lists thirteen parameters: besides performative, sender, receiver, reply-to and content there are language, encoding and ontology, which tell the recipient in which formalism and with which vocabulary the content must be read, plus five conversation control parameters (protocol, conversation-id, reply-with, in-reply-to, reply-by). The Communicative Act Library defines twenty-two of them, from inform and request through call for proposal, accept-proposal and subscribe. The detail that matters most is how little compliance actually demands: the only mandatory parameter is the performative, and the only act every agent must implement is not-understood. The most widely used implementations are JADE, in Java, and SPADE, in Python over XMPP.

An enterprise example

A manufacturing group has to place a rush order across four plants with different loads and capabilities. In the Contract Net Interaction Protocol (SC00029) the agent representing the commercial function sends a call for proposal to the four plant agents; each answers with a propose carrying a date and a cost, or with a refuse when it has no capacity. The initiator sends an accept-proposal to the winner and a reject-proposal to the others, and the winner closes with inform-done or failure. It is the same shape found today in an agent swarm with an orchestrator and workers, except it was written more than twenty years earlier, with the semantics of every step pinned down in a specification.

Why it matters for decision makers

FIPA's own site, where the organization stated that it was no longer active, is gone: the domain has been reassigned, and that sentence now reads only in the archives, where the specifications live too. It is worth understanding why so complete a standard saw such thin industrial adoption. Its formal semantics assumed agents able to reason about each other's beliefs and intentions, a capability the software of the day did not have; and compliance that settles for a single mandatory act guarantees the ability to speak, not to understand. Anyone assessing A2A or MCP today faces the same question with the terms reversed: language models finally make intent interpretation practical, while the layer that was well standardized back then, act semantics and conversation protocols, is the part still to be built. Shared vocabulary was not solved by FIPA either, which only provided a field in which to declare it.