Overview

As per the course website, "This week’s lecturecast focuses on the way in which agents communicate with each other. The lecturecast also looks at the theory of language with Speech Acts and then provides an example of an agent communication language (KQML). Finally, an introduction to ontologies and the way they can be used in agent communications will be discussed."

My Reflection

Overall Reflection

This week was an interesting one. It focused on communication between agents, but to do so it started from the philosophy of language between humans, which was interesting and satisfying for me.

The lecturecast of the unit gave an overview for both topics, introducing Speech Acts that was pioneered by J. L. Austin and John R. Searle, and is a theory that is generally concerned with how utterance of words initiates actions in a social context, rather than being mere expression of a proposition. The lecturecast then proceeded to introduce Agent Communication Languages (ACLs), espcially Knowledge Query and Manipulation Language (KQML) and its Knowledge Interchange Format (KIF) that can be used to express ontologies.

The readings then covered the same area with two reading items. The first was the introduction and Chapter 1 from Speech Acts book by Searle (1969). The second was the paper Negotiating over Ontological Correspondences with Asymmetric and Incomplete Knowledge by Payne and Tamma (2014).

Readings Reflection

It was interesting for me to read about philosophy of language through Speech Acts, from which I learned about locutionary, illocutionary and perlocutionary acts. Below are some of the quotes that I liked from the introduction and the first chapter:

• "I take it to be an analytic truth about language that whatever can be meant can be said. A given language may not have a syntax or a vocabulary rich enough for me to say what I mean in thatlanguage but there are no barriers in principle to supplementing the impoverished language or saying what I mean in a richer one."
• "The hypothesis then of this work is that speaking a language is engaging in a rule-governed form of behavior. To put it more briskly, talking is performing acts according to rules."
• "[S]ome philosophers ask, 'How can making a promise create an obligation?' A similar question would be, 'How can scoring a touchdown create six points ? ' As they stand both questions can only be answered by citing a rule of the form, ' X counts as Y', which is, of course, not to say that the questions cannot be rephrased to ask important questions about the institution of promising—or for that matter, football."
• "Sometimes in order to explain adequately a piece of human behavior we have to suppose that it was done in accordance with a rule, even though the agent himself may not be able to state the rule and may not even be conscious of the fact that he is acting in accordance with the rule."
• "In an article entitled Meaning^ Grice gives the following analysis of the notion of "non-natural meaning". To say that a speaker S meant something by X is to say that S intended the utterance of X to produce some effect in a hearer H by means of the recognition of this intention."

For the second reading on Negotiating over Ontological Correspondences with Asymmetric and Incomplete Knowledge, the paper aimed to introduce the concept of Correspondence Inclusion Dialogue (CID), a protocol of multi-agent communication, by which agents negotiate by exchancing beliefs of the utilities for each correspondence. Published in 2014, the paper introduced CID protocol to solve the gap of having prior research concerened with asymmetric-knowledge agents communication focusing on two approaches. The first was that agents exchange messages of conceptual definitions so that each agent would mutate their ontologies for alignment, and the second was to focus on argumentation techniques on the type of correspondence to discover mutully acceptable alignments. Both approaches rely on having the onotologies of the agent parties being fully shared. Instead, the CID that was proposed by this paper is built on:

1. "Each agent is aware of a set of correspondences, each with an associated degree of belief."
2. "There should be no ambiguity with respect to either the source entities in the resulting alignment, or the target entities."
3. "If alternative choices of correspondences exist, the selection is based on the combined, or joint degree of belief of both agents."
4. "No correspondences should be selected where their joint degree of belief is less than some defined admissibility threshold."
5. "The alignment should be generated by disclosing as few beliefs as possible."

The paper then proceeds to give an example implemetation between two agents, named Alice and Bob, with illustrative tables and figures for the correspondences and negotionation sequence and associated degrees of belief and how they map between ontological entities.

While reading the paper, I had two main concerns. One being that I obviously do not know how I can practically implement this. The second was that I had concern that the paper is from 2014, and I kept thinking how it relates to the present, and where CID falls among the recent agent communication protocols. I recently learnt about the Agent 2 Agent (A2A) communication protocol, and at the time of writing this page, I have a challenge at work to develop an LLM-based agent using A2A protocol, so I am highly focused on my need to learn more about it.

To tackle the second concern at least, I did a quick research from which I learnt that CID and A2A are from very different eras, assumptions and technical motivations, and the relationship can be understood as conceptual lineage. While CID is an early, formal and logic-based negotiation protocol that was designed for symbolic agents to exchange beliefes about ontology correspondences, A2A is by contrast a modern, LLM-driven communication paradigm, in which agents coordinae using natural language, implicit reasoning, and learned representations.

I then started thinking that maybe logic-based protocols, like CID can be integrated with A2A to provide a semantic layer to agent communications, constructing a guardrail for multi-agent syestems, and idea that I had from the previous Knowledge Representation and Reasoning module, and that I have pointed to in one of my peer responses on the first collaborative discussion during Unit 2. I looked up any research on possible integration between CID and A2A, and found none that is specifically on the topic. However, I found an interesting paper on the use of logic formalisation on top of a multi-agent system (MAS) built for cross-domain art translation (Matanski et al., 2025), which was also interesting for given it tackles the use for intelligent agents in the artistic practices, a domain which I intend to pursue after my master's degree. To sum up, I'm still not sure whether the lack of reasearch on possible integration between CID and A2A is a reasearch gap or a question that has become simply outdated, but that in itself is a question to be further explored.

Matanski, V. et al. (2025) ‘Ontology-Driven Multi-Agent System for Cross-Domain Art Translation’, Future Internet, 17(11), pp. 517–517. Available at: https://doi.org/10.3390/fi17110517.

Payne, T.R. and Tamma, V. (2014) ‘Negotiating over Ontological Correspondences with Asymmetric and Incomplete Knowledge’, in AAMAS ’14: Proceedings of the 2014 international conference on Autonomous agents and multi-agent systems. Richland, SC: International Foundation for Autonomous Agents and Multiagent Systems, pp. 517–524. Available at: https://dl.acm.org/doi/10.5555/2615731.2615816 (Accessed: 28 February 2026).

Searle, J.R. (1969) Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press, pp. 1–53.