LLM Agent Communication Protocol (LACP) Requires Urgent Standardization: A Telecom-Inspired Protocol is Necessary¶
Conference: NeurIPS 2025 arXiv: 2510.13821 Code: To be confirmed Area: LLM Agent Keywords: multi-agent communication, protocol standardization, agent interoperability, telecom-inspired design, security-by-design
TL;DR¶
This position paper argues that the fragmented ecosystem of current LLM Agent communication mirrors the "protocol wars" of the early networking era. It proposes LACP, a three-layer protocol (Semantic, Transactional, and Transport layers) inspired by telecom standardization, and contends that security-by-design, transactional integrity, and semantic interoperability are critical for multi-agent systems.
Background & Motivation¶
Current LLM Agent communication suffers from severe fragmentation, with numerous protocols operating in isolation:
| Protocol | Release Date | Developer | Interface Type | Security Features |
|---|---|---|---|---|
| Function Calling | 2023.06 | OpenAI | JSON schema | API key only |
| Agent Protocol | 2024.11 | LangChain | REST API | HTTP/JWT |
| MCP | 2024.11 | Anthropic | JSON-RPC/HTTP | OAuth 2.1 |
| ACP | 2024 | IBM/LF | JSON-RPC | Signed token, RBAC |
| ANP | 2024.03 | Community | DID/JSON-LD | W3C DIDs |
| Agora | 2024.10 | Academia | Meta-protocol | Hash-based ID |
| A2A | 2025.04 | HTTP/Protobuf | Capability discovery |
This fragmentation introduces three systemic risks:
Lack of Interoperability: Agents across different frameworks cannot collaborate seamlessly, making heterogeneous multi-agent system development difficult.
Security as an Afterthought: Security is not a core component, leaving systems vulnerable to data tampering, agent impersonation, and related attacks.
Monolithic Design and Absence of Transactional Integrity: Communication logic is tightly coupled with agent implementation, and complex multi-step operations lack atomicity guarantees.
The authors draw a historical analogy: this situation mirrors the "protocol wars" of the 1970s–1990s, which were only resolved by the widespread adoption of TCP/IP. Without a shared communication substrate, the transformative potential of distributed AI cannot be realized.
Method¶
Overall Architecture¶
LACP is a three-layer protocol architecture inspired by the layered abstraction principles of telecommunications engineering:
┌─────────────────────────────┐
│ Semantic Layer │ ← Conveys communication intent
├─────────────────────────────┤
│ Transactional Layer │ ← Ensures reliability and integrity
├─────────────────────────────┤
│ Transport Layer │ ← Efficient and secure message delivery
└─────────────────────────────┘
The three layers are mutually isolated with clearly defined interfaces and can evolve independently.
Key Designs¶
Semantic Layer:
Defines a minimal set of common message types following a "narrow waist" design principle:
| Message Type | Required Fields | Optional Fields | Purpose |
|---|---|---|---|
| PLAN | intent_id, role, natural_language | graph_ops | Express high-level intent |
| ACT | intent_id, tool_call, params | deadline, cost_cap | Invoke external tools |
| OBSERVE | intent_id, status, output | metrics | Return results/status |
All messages are wrapped in a JWS (JSON Web Signature) envelope to ensure semantic security and clarity.
Transactional Layer:
Provides the following mechanisms to ensure communication reliability: - Message Signing: End-to-end cryptographic integrity verification - Message Ordering: Ensures sequential consistency for multi-step operations - Unique Transaction IDs: Guarantees idempotency and prevents replay attacks - Atomic Transaction Support: Draws on the concept of two-phase commit (2PC)
Transport Layer:
Transport-agnostic design, operable over multiple network protocols: - HTTP/2 - QUIC - WebSockets
This enables LACP to adapt to diverse network environments, including next-generation networks such as 6G.
Design Principles (Insights from Telecom Standardization)¶
- Consensus-Driven Open Standards: Modeled on the collaborative approach of ITU and 3GPP
- Security by Construction: Security is not an add-on but a foundational component at every layer
- Layered Abstraction: Inspired by the OSI model with clear separation of concerns
- Narrow Waist Principle: Minimal core with extensible edges, analogous to the IP protocol
Key Experimental Results¶
Main Results¶
Performance and Overhead Analysis (10,000 sequential requests):
| Scenario | Baseline Size | LACP Size | Size Overhead | Baseline Latency | LACP Latency | Latency Overhead |
|---|---|---|---|---|---|---|
| Small (51B) | 51 bytes | 306 bytes | +500% | 0.85 ms | 0.88 ms | +3.5% |
| Medium (151B) | 151 bytes | 442 bytes | +191% | 0.86 ms | 0.89 ms | +3.1% |
| Large (1964B) | 1,964 bytes | 2,560 bytes | +30% | 0.89 ms | 0.92 ms | +2.9% |
Key findings: Latency overhead is negligible (only +0.03 ms for large payloads); payload overhead for practically sized messages is approximately 30%.
Interoperability Validation:
Transparent communication between a LangChain ReAct Agent and a standalone LACP Tool Server was successfully demonstrated: - The LangChain Agent constructs and sends a signed LACP ACT message - The LACP server verifies, executes, and returns a signed OBSERVE message - No framework-specific custom integration code is required throughout
Ablation Study¶
Security Validation — Tampering Attack: - A valid signed message is generated (transfer amount: 100) - The amount is altered to 10,000 after signing - Server-side signature verification immediately fails, returning HTTP 403
Security Validation — Replay Attack:
- A legitimate message is sent and successfully processed
- The same message is retransmitted
- Signature verification passes, but the transactional layer detects a duplicate transaction_id and returns HTTP 409
Key Findings¶
- LACP latency overhead is negligible: For practical payloads, the absolute latency increase is only 0.03 ms
- End-to-end security is not equivalent to transport-layer security: TLS only protects data in transit and cannot prevent tampering or replay attacks at endpoints; LACP's application-layer security is a necessary complement
- The PLAN/ACT/OBSERVE semantic model is compatible with existing frameworks: Successfully integrated into LangChain workflows
Highlights & Insights¶
- The telecom analogy is apt: The parallel between agent communication fragmentation and the networking "protocol wars" is compelling and well-argued
- The three-layer architecture is well-motivated: The separation of semantic, transactional, and transport concerns is clean, with clearly defined responsibilities at each layer
- Security by design rather than as an afterthought: This is critical for high-stakes multi-agent applications in finance, healthcare, and autonomous driving
- "Narrow waist" design philosophy: The three core message types—PLAN, ACT, and OBSERVE—are sufficiently concise yet expressive
- Validated with a working prototype: The paper goes beyond conceptual discussion by providing a functional code implementation and quantitative evaluation
Limitations & Future Work¶
- Position paper only: Large-scale deployment validation in real-world multi-agent systems is absent
- Limited performance testing scope: Experiments were conducted locally on a single machine; cross-regional and high-concurrency scenarios were not evaluated
- No incremental migration path from existing protocols: Protocol replacement carries high practical costs and requires bridging strategies
- Expressive power of the semantic layer remains to be verified: Whether PLAN/ACT/OBSERVE covers all agent interaction patterns (e.g., negotiation, bidding) is unclear
- Competition with MCP and A2A: Google A2A and Anthropic MCP already have industry backing; the standards competition is unresolved
- Lack of dynamic discovery and registration mechanisms: How agents discover and register within an LACP network is insufficiently discussed
- Cryptographic overhead in high-frequency scenarios: Per-message JWS signing may become a bottleneck in high-frequency, short-message agent communication
Related Work & Insights¶
- vs. MCP: MCP focuses on standardizing tool-resource-prompt context, whereas LACP addresses broader agent-to-agent communication security and interoperability
- vs. A2A: A2A supports peer-to-peer communication and Agent Cards but lacks atomicity guarantees at the transactional layer
- vs. ANP: ANP employs W3C DIDs for decentralized identity; LACP provides a more complete layered security model
- Inspiration from the OSI seven-layer model: LACP simplifies to three layers, better suited to the practical demands of agent communication
- Insight: A window for standardizing multi-agent systems genuinely exists; establishing standards early is far less costly than remediation after fragmentation has entrenched itself
Rating¶
- Novelty: ⭐⭐⭐ The telecom-inspired agent protocol design is meaningful, though the core architectural ideas are not entirely novel
- Experimental Thoroughness: ⭐⭐⭐ A prototype experiment is provided but limited in scale; acceptable for a position paper
- Writing Quality: ⭐⭐⭐⭐ The argumentation is well-structured, historical analogies are vivid, and technical descriptions are accurate
- Value: ⭐⭐⭐⭐ Raises an important standardization call at a critical juncture in the agent protocol landscape