Structured Inter-Agent Communication

• Structured inter-agent communication is a formal framework that defines protocols, message schemas, and state machines to coordinate multi-agent interactions.
• It incorporates methodologies like FSMs, graph-based models, and attention mechanisms to ensure secure, consistent, and efficient information exchange.
• Applications span decentralized control, multi-agent reinforcement learning, robotics, and semantic web integration, demonstrating practical robustness in diverse environments.

Structured inter-agent communication defines the formal mechanisms, protocols, and architectures by which autonomous agents exchange information, negotiate actions, and coordinate behaviors in multi-agent systems (MAS). In contrast to gossip-based or purely ad hoc message passing, structured approaches impose syntactic and semantic constraints on messages, specify interaction patterns, manage protocol state, and enforce security and consistency. These frameworks are foundational for decentralized control, collaborative problem solving, multi-agent reinforcement learning (MARL), and enterprise integration, with applications spanning robotics, LLM-driven agentic AI, distributed control systems, and the Internet of Agents.

1. Formal Protocols and State-Machine Models

Structured communication protocols in MAS are formally modeled as finite-state machines (FSMs), state-transition systems (STS), or Petri nets, providing explicit specification of message types, sequencing, roles, and valid actions. Protocols such as FIPA-ACL, BPEL4WS, AUML, Agent Communication Protocol (ACP), Model Context Protocol (MCP), and Agent-to-Agent (A2A) define envelopes, performatives, and schema for agent messaging (Benmerzoug, 2013, Rizvi-Martel et al., 14 Oct 2025, Liu et al., 18 May 2025, Ehtesham et al., 4 May 2025, Hu et al., 2024). Each protocol can be described as a tuple $P = (\mathcal{S}, \mathcal{M}, \mathcal{T}, s_0)$ where:

• $\mathcal{S}$: Finite set of protocol states.
• $\mathcal{M}$: Typed messages (including speech-act performatives: REQUEST, INFORM, AGREE, REFUSE, PROPOSE, etc.).
• $\mathcal{T}$: State transitions $(s, m, s')$ labeled by roles and message conditions.
• $s_0$: Initial protocol state.

Message payloads are validated against machine-readable schemas (JSON-Schema, Protobuf, RDF/OWL). Protocol enforcement is achieved by agent-side state machines that advance state upon valid message receipt or issuance, guaranteeing ordering, role compliance, and error handling. For example, Benmerzoug et al. describe XML-based BPEL interaction protocols for purchase orders; ACRE implements FSM-based conversation management across agent platforms, with XML exchange, automatic variable binding, and runtime event triggers (Benmerzoug, 2013, Lillis, 2017).

2. Graph-Based and Attention-Driven Communication Structures

Recent advances in MARL and agentic AI leverage graph-centric and differentiable architectures for dynamic, scalable inter-agent communication. Frameworks such as DMCA, CommFormer, and MACTAS represent agent interactions as learnable, sparsified graphs, attention matrices, or cascaded pipelines (Arul et al., 2022, Hu et al., 2024, Wojtala et al., 19 Aug 2025, Meneghetti et al., 2020, Wang et al., 21 Jun 2025).

• Learnable Communication Graphs: CommFormer formulates communication as bi-level optimization over a real-valued adjacency matrix $\alpha \in \mathbb{R}^{N \times N}$, relaxed via k-hot Gumbel-Max sampling and dynamically gated per agent and timestep. Messages are routed, attended, and fused only along the selected links, with task-dependent sparsity (Hu et al., 2024).
• Self-Attention-Based Communication: DMCA and MACTAS employ multi-head attention (query/key/value projections) over sets of agent states or hidden vectors, yielding structured, weighted fusion of information across neighbors. DMCA further learns selective communication links via link-prediction MLPs and aggregates messages via LSTMs (Arul et al., 2022, Wojtala et al., 19 Aug 2025). Specialized relational GCNs are used for class-wise message transforms in heterogeneous MAS (Meneghetti et al., 2020).
• Sequential Adaptive Routing: AnyMAC reframes agent collaboration as sequential next-agent prediction and context selection, expanding communication topology search space beyond fixed graphs (Wang et al., 21 Jun 2025).

3. Protocol Suites for the Internet of Agents (IoA)

A new generation of protocol stacks for IoA integrates layered and extensible communication, semantic negotiation, and capability orchestration (Fleming et al., 24 Nov 2025, Liu et al., 18 May 2025, Ehtesham et al., 4 May 2025, Ren et al., 23 Aug 2025):

• Layered Stack: The IoA stack adds Layer 8 (Agent Communication Layer, L8) for standardized message envelopes and speech acts, and Layer 9 (Agent Semantic Negotiation Layer, L9) for shared context negotiation and schema locking, above HTTP/2–3 (Fleming et al., 24 Nov 2025). L8 formalizes interaction patterns; L9 ensures semantic grounding and compatibility for all content.
• Agent Collaboration Protocols (ACPs): The ACPs suite defines ARP (Registration), ADP (Discovery), AIP (Interaction/Workflows), and ATP (Tooling) as FSMs, with typed message schemas, explicit state transitions, ordered handshakes, heterogenous agent support, and semantic modeling via JSON-LD and OWL ontologies (Liu et al., 18 May 2025). Example applications orchestrate workflows such as restaurant booking via sequence diagrams and state machines.
• MCP/A2A/ANP/ACP Survey: The comparative survey details key dimensions—client/server, RESTful brokered, peer-to-peer, decentralized; JSON-RPC, multipart MIME, Agent Cards, DID/JSON-LD; registry-based or decentralized discovery; authentication and schema-based security (Ehtesham et al., 4 May 2025). Anemoi demonstrates semi-centralized thread-based coordination atop A2A-MCP, minimizing redundant context passing and supporting plan refinement, consensus, and real-time bottleneck detection (Ren et al., 23 Aug 2025).

4. Expressivity, Bandwidth, and Communication Complexity

Theoretical analyses map structured inter-agent communication regimes to labeled DAGs, associating expressivity, speedup, and trade-offs (Rizvi-Martel et al., 14 Oct 2025):

• Associative Recall: Simple key–value lookup achieves $O(1)$ depth and communication with $\Theta(N)$ agents.
• Parallel State Tracking: Prefix-sum and parity tasks admit scalable parallelism at depth $D(N,w)=O(N/w+\log w)$ and communication $B(N,w)=O(w)$.
• $k$-Hop Reasoning: Nested queries require $O(k)$ wall-clock depth and $O(k)$ communication, unavoidably scaling linearly in reasoning depth.
• Optimal protocols enforce single-token message formats (SEND, BROADCAST, RECEIVE) and explicit agent–time DAG connectivity, matching theoretical bounds in LLM experiments.

5. Semantic Web Integration and Protocol Relationship Discovery

Semantic Web frameworks formalize communication acts as OWL/DL classes and branches of protocols as state–fluent traces, enabling cross-protocol comparison and interoperability (Berges et al., 2024):

• Ontology-Driven Messaging: Communication acts are subsumed under CommOnt ontology; acts such as Request, Accept, Responsive inherit formal relationships and semantics.
• Event Calculus Semantics: Social commitments (C, CC) are represented as fluents, with Event Calculus rules for initiation, termination, and holding.
• Protocol Relationship Discovery: Semantic relationships (equivalence, specialization, prefix, suffix, infix, complement) are algorithmically inferred by comparing sets of generated fluents per protocol branch, supporting interoperability and cross-domain mapping.

6. Gossip Protocols and Hybrid Communication Substrates

Gossip (epidemic) protocols provide a decentralized, fault-tolerant diffusion layer complementary to structured RPC/tool invocation (Khan et al., 2 Dec 2025, Habiba et al., 3 Aug 2025):

• Performance Models: Under push–pull with fan-out $f$, convergence time is $O\left(\frac{\ln N}{\ln(1+f)}\right)$; staleness decays exponentially per round; random peer-sampling (e.g., CYCLON, SCAMP) preserves connectivity under churn up to $O(k)$.
• Semantic Filtering and Knowledge Decay: Adaptive relevance scoring, TTL/priority tags, trust propagation, and spectral analysis manage message relevance and temporal staleness.
• Hybrid Architectures: Structured protocols (for transactions, negotiations) operate in tandem with gossip (for discovery, consistency, robustness) via agent-side dual layers, anti-entropy reconciliation, and semantic prioritization.

7. Interpretable Emergent Language and Symbolic Protocols

Recent transformer-based models in MARL demonstrate the emergence of interpretable symbolic languages for agent communication (Bhardwaj, 4 May 2025):

• Discrete Message Vocabularies: Differentiable inter-agent transformers (DIAT) jointly train speaker/listener networks with hard argmax symbol selection, yielding messages clustering into meaningful, human-readable alphabets encoding environmental states.
• Compositionality and Hierarchy: Emergent codes exhibit compositional and hierarchical structure, allocating symbol classes or token positions to encode color, shape, spatial layout, or macro-roles.

Table: Families of Structured Communication Protocols in MAS

Protocol Name	Architecture	Typical Use Case
FIPA-ACL	FSM, Sch./AUML	General agent messaging
MCP	JSON-RPC, Med/Broker	Tool invocation, LLM control
ACP	REST, FSM	Capability registry, workflows
A2A	Peer RPC/SSE	Enterprise agent delegation
CommFormer	Diff. graph, Transformer	Learned coordination graphs (MARL)
DMCA/MACTAS	Attention, GNN	Decentralized navigation, RL
IoA L8/L9	Layered, Schema	Semantic agent negotiation
ACRE	FSM, XML	Platform conversation management

Structured inter-agent communication, via FSMs, layered protocols, learned graphs, and ontological semantics, underpins scalable, reliable, and semantically meaningful coordination across heterogeneous agent ecosystems. Ongoing work addresses the challenges of bandwidth efficiency, dynamic topology, semantic interoperability, resilience under churn, emergent clustering, and hybrid integration with decentralized gossip substrates.