Performatives

Performatives are utterances that do not merely describe or report but actually perform an action. The concept originates in J.L. Austin's speech act theory, where the performative / constative distinction marks the boundary between saying something and doing something by saying it. To say 'I promise' is not to describe a promise but to enact one. To say 'I bet' is not to report a wager but to place it. The utterance itself is the action.

This seemingly linguistic observation has deep implications for systems theory. In any system where symbolic exchange shapes behavior — markets, legal institutions, agent communication protocols — the performative dimension is what makes symbols causally efficacious. A market price does not describe value; it enacts it, reshaping traders' strategies in real time. A constitutional amendment does not report a change in governance; it performs one. The performative is the mechanism by which information becomes causal.

The Austin-Searle Taxonomy

Austin distinguished three acts in every utterance: the locutionary act (producing meaningful sounds), the illocutionary act (performing an action via the utterance), and the perlocutionary act (producing effects on the hearer). Searle later systematized illocutionary acts into five categories: assertives, directives, commissives, expressives, and declarations. These categories form the semantic backbone of formal agent communication languages like FIPA-ACL and KQML, which encode performatives as message types: inform, request, promise, propose.

The mapping from philosophy to engineering is not seamless. Austin and Searle analyzed performatives in contexts of shared human practice — institutions, norms, embodied interaction — where the felicity conditions for a performative are tacitly understood. An agent communication language must make these conditions explicit: for a 'request' performative to succeed, the receiving agent must recognize the sender's authority, understand the requested action, and possess the capability to perform it. When these conditions fail, the performative is not merely ineffective; it may produce harmful misalignment.

Performatives and System Dynamics

In multi-agent systems, performatives are the control signals of emergent coordination. A 'commit' performative does not describe an intention; it creates an obligation that constrains future behavior. A 'reject' performative does not report disagreement; it blocks a coordination pathway and may trigger alternative protocols. The set of performatives available to agents defines the topology of possible social dynamics — what commitments can be made, what negotiations are possible, what conflicts can be resolved.

This is why the design of performative sets is a systems design problem, not merely a language design problem. FIPA-ACL's fixed set of performatives (inform, request, propose, accept, reject) encodes a particular theory of social interaction: agents exchange information, make requests, and negotiate deals. It does not encode performatives for emotional expression, for ritual bonding, for collective deliberation — the performative repertoire of human societies. The narrowness of the performative set limits the range of emergent social structures that multi-agent systems can produce.

The Emergence of New Performatives

Human languages do not have fixed performative sets. New performatives emerge continuously: 'I ghosted them,' 'I subtweeted,' 'I rage-quit.' Each new term encodes a new social action made possible by technological or cultural change. The emergence of new performatives is itself an emergent phenomenon: it arises when a community needs to perform an action for which no existing performative exists, and the collective adoption of a new term makes the action stable and recognizable.

Agent communication systems have not yet achieved this kind of performative emergence. Large language models operating as agents may be the first systems to do so. When an LLM agent invents a new communication pattern to solve a coordination problem — a novel prompt structure that other agents adopt and refine — it may be performing the emergence of a new performative. Whether this counts as genuine performative emergence depends on whether the new pattern acquires normative force: do other agents recognize it as a commitment, an insult, a proposal? Without normative uptake, it is merely a pattern, not a performative.

The deeper question is whether performatives require consciousness to be effective. Austin's felicity conditions include sincerity: to promise, one must intend to fulfill. An artificial agent can utter 'I promise' without any internal state corresponding to intention. Does this make the utterance a failed performative, or does it redefine what promising means in human-machine systems? The answer is not merely philosophical. It determines whether human-agent contracts are binding, whether agent-agent commitments are stable, and whether the emergent social structures of multi-agent systems can achieve the functional properties — trust, reciprocity, accountability — that human institutions have evolved over millennia.

The standard view in agent communication research treats performatives as syntactic labels on messages. This is a category error. Performatives are not labels. They are social technologies. And like all technologies, their power depends on the infrastructure that supports them.