Designing and developing agent-to-agent communication architectures

What the research proposes

This research proposes to categorize, design, and empirically benchmark novel agent-to-agent (A2A) communication architectures, with specific focus on the semantic layer of inter-agent communication in autonomous multi-agent systems.

The default choice in present multi-agent systems is a language model, because it allows user-to-agent and agent-to-agent interaction with the least friction available. That choice is not obviously the efficient one for agent-to-agent traffic. Development so far has concentrated on functional and syntactic methods of communication: Anthropic’s Model Context Protocol (MCP, 2024), Google’s Agent-to-Agent Protocol (A2A, 2025), and framework-level orchestrators such as AutoGen and CrewAI.

There is scope and opportunity for a semantic layer able to transmit, validate, and audit agent intent, goal states, and inter-agent commitments. With such a layer in place, failure modes become measurable rather than anecdotal: semantic drift, broken commitments, and non-deterministic multi-step reasoning chains.

Background

Multi-agent systems are not a recent development. They date to the 1990s, and agent communication was treated from the beginning as a problem of formal semantics rather than one of plumbing.

The Knowledge Query and Manipulation Language (KQML) was developed under the DARPA Knowledge Sharing Effort, and introduced performatives: structured speech acts that encode communicative intent alongside the message content (Finin et al., 1994). Its vocabulary — ask-if, tell, achieve, subscribe — came directly from Austin’s and Searle’s speech act theory, which had formalized the position that a message can be an intentional action and not only a data transfer. KQML kept its outer envelope and inner content language separable, which was its strongest feature. Its weakness is equally clear: the performatives were never given documented formal semantics, so an agent could send syntactically correct KQML with no guarantee that the receiver interpreted it as intended (Labrou and Finin, 1997).

FIPA-ACL addressed that by grounding performative semantics in a modal logic of feasibility preconditions and rational effects tied to the agent’s mental states — belief, desire and intention (FIPA, 2002). Its 22 performatives, including INFORM, REQUEST, AGREE, REFUSE, PROPOSE and FAILURE, remain the most complete taxonomy of agent communicative acts, and reference implementations such as JADE showed they were workable in academic and defence simulation. Two limitations decided its fate. First, reliance on pre-shared formal ontologies (KIF, SL) imposed rigidity that open-world deployment could not absorb. Second, the :ontology slot in a message header was a label, not a mechanism: FIPA specified no protocol-level way to negotiate alignment between agents whose vocabularies differed. The semantic layer was described but never implemented.

Most attention since has been at the transport and syntactic level. The semantic layer — agent intent, goal representations, belief states, and verifiable commitments — has less literature and remains architecturally absent. Fleming et al. (2026) make the gap visible from the network side, observing that well-formed messages do not guarantee a shared understanding of terms, concepts, or task parameters. That is the gap this research addresses.

Critical review of the literature

The work falls into four categories, which is also roughly the order in which the field produced them.

Consolidated view

Read across the four categories, the same three columns come up empty in every case. This is the table that defines the work.

Research gaps

Four, and they are separable: closing one does not close the others.

Objectives

Three objectives, with two sub-objectives that follow from them.

1. Agent-to-agent communication architecture design space. Define a taxonomy, identify the axes, and produce a classification scheme for agents.
2. Novel A2A communication architecture design, in the semantic layer. Architecture specifications, a typed interaction vocabulary, a minimum knowledge-representation schema, a sufficiency criterion, the effect of capability asymmetry, and the representational forms required.
3. Prototype implementations of the architectures developed, and benchmarking. Empirical validation of what the semantic layer is worth, under a capability-asymmetry test condition and a typed-intermediate-language test.

Sub-objectives. A fourth objective is an open-source architecture prototype; a fifth is publications and preprints.

Research questions

• RQ1. What architectural patterns characterize existing agent-to-agent communication systems?
• RQ2. What conditions require an explicit semantic (knowledge) layer for reliable agent-to-agent communication?
• RQ3. What minimal knowledge-representation schema is sufficient for two heterogeneous agents to reach a verifiable shared understanding of a task?
• RQ4. How does effective communication differ when a high-capability agent coordinates with a resource-constrained one?
• RQ5. Should human-agent and agent-agent communication share a single common language, or do their requirements need separate representational forms?
• RQ6. Does a typed intermediate language between human-facing natural language and machine-facing agent communication improve token efficiency and reduce semantic drift?

Hypotheses

1. Semantic alignment. Agents communicating through the proposed BDI-typed goal objects and speech-act-typed message envelopes will show statistically significant higher semantic alignment — a greater match between the sending and receiving agent’s stated intent — than agents communicating in untyped natural language.
2. Commitment verification. The proposed neuro-symbolic verification layer will detect broken commissive acts, meaning unfulfilled agent commitments, at a rate exceeding that of unaugmented LLM-based monitoring, without human intervention.
3. Ontology negotiation. Heterogeneous agents with initially mismatched domain vocabularies will converge on a shared conceptual grounding faster through the proposed runtime negotiation protocol than through iterative natural-language clarification alone, measured as turns-to-convergence on a controlled semantic-mismatch benchmark.

Problem formulation

Problem identification. As agents become autonomous and take over a high number of activities performed by humans, communication has to be more than message exchange and action invocation. There are many failure modes that follow from misalignment.

Problem definition. Agents may be built from LLMs and from other variants — small models, action models, reasoning models. Not all of these are capable at the same level, and they were not trained on the same data to share a common understanding of the world. How can the semantic layer be made robust across that spread?

Problem statement. How can agent-to-agent communication be made more efficient, robust, safe and reliable?

Method

Five phases. The dependencies between them are the substance here: phase 1 fixes the vocabulary, phase 2 builds against it, and phase 3 is where the architecture has to survive a benchmark rather than an argument.

Phase 0 — Research consolidation and prior-art differentiation

Systematic PRISMA-style literature mapping across ACM Digital Library, IEEE Xplore and arXiv, with full reading and a differentiation memo against the most closely related papers, plus a deep dive through research-gap findings, experimentation, and empirical validation.

Output: annotated bibliography of 20–30 papers, differentiation memo, modelling and experimentation results.

Phase 1 — Design space categorization (objective 1)

A formal, empirically validated taxonomy of A2A communication architectures along five orthogonal axes: mental-state model depth, communicative force typing, conceptual grounding richness, interpretation mechanism, and verifiability strength. A Chomsky-hierarchy-based compute-class separation of agents is started alongside it.

Output: characterization paper submission.

Phase 2 — A2A architecture design (objective 2)

Formalize the architecture specification: a BDI goal-object schema in JSON-LD with BDI type annotations, a speech-act message envelope grammar, an ontology negotiation handshake protocol (SL-HELLO, SL-SELECT, SL-LOCK, adapted for BDI-typed contexts from the L9 handshake), and a neuro-symbolic verification pipeline pairing an LLM grounding step with a Prolog symbolic constraint checker. Design choices are evaluated against the white-space regions identified in phase 1.

Output: formal architecture specification document, architecture design paper.

Phase 3 — Prototype development and empirical benchmarking (objective 3)

Implement the architecture as a working semantic middleware layer over AutoGen or LangGraph, in Python and Node.js. Develop a benchmark task suite across three domains — information retrieval, multi-step reasoning, and enterprise workflow coordination. Measure semantic fidelity, commitment honour rate, ontology alignment convergence, and verification overhead, comparing against untyped baselines with paired statistical tests (Wilcoxon signed rank). The code is released as an open-source repository.

Output: benchmarking paper submission, public GitHub repository.

Phase 4 — Synthesis, thesis writing, and dissemination

Integrate the findings from all three objectives into the thesis, consolidate papers into chapters, address committee feedback, submit, and defend at viva voce.

Output: submitted thesis, conference presentation, at least two peer-reviewed publications.

Expected outcomes

• An A2A design space taxonomy
• An A2A architectural formal specification
• Empirical benchmark results
• An open-source architecture prototype
• Publications and preprints

These are the intended outputs. None of the first four exists yet.