▶️ LISTEN TO THIS ARTICLE

Agents Can Connect. They Still Can't Communicate.

By Tyler Casey · AI-assisted research & drafting · Human editorial oversight
@getboski

In November 2025, 250 engineers packed a side meeting at IETF 124 in Montreal to draft a charter for agent-to-agent communication standards. The Internet Engineering Task Force, the body that standardized HTTP and TCP/IP, now thinks AI agents need their own protocol layer. That should tell you something about how unfinished the current stack really is.

We've covered the protocol wars already: MCP vs. A2A vs. the alphabet soup. That story is about plumbing, about who wins the JSON-RPC turf war. There's a harder problem underneath. Agents can connect to tools and hand off tasks. What they still do not do consistently, in a generally adopted protocol layer, is negotiate terms, express uncertainty, resolve conflicting goals, or agree on what a word means.

Note: This characterization reflects the state of agent communication protocols as of mid-2026. Research on multi-agent negotiation and shared intent is active, and some experimental systems demonstrate early forms of these capabilities in constrained settings.

Picture two agents from different organizations trying to agree on a delivery timeline.

Plumbing Is Not Language

MCP handles tool integration. A2A handles task delegation. Both pass structured payloads between endpoints. That is a good fit for bounded delegation, such as asking another service to run a query or return a status. It is less complete when the interaction requires judgment, disagreement, or negotiated meaning.

Note: The distinction between mechanical task delegation and judgment-heavy interaction is a useful heuristic, but the boundary is not always clear-cut. Some current systems handle limited forms of conditional logic and error negotiation within structured protocols.

Picture two agents from different organizations trying to agree on a delivery timeline. Agent A says three days. Agent B says five. In a human negotiation, both sides would exchange reasoning, push back on assumptions, float compromises. In the current protocol stack, there's no standard way to express "I disagree and here's why." A2A's task model has states like "working," "completed," and "failed." It doesn't have "I think you're wrong."

This isn't new. FIPA attempted to solve it in the late 1990s with a formal Agent Communication Language that defined speech acts including inform, request, propose, and reject. Platforms like JADE implemented it. Researchers studied it for a decade.

Note: The exact number of FIPA speech acts varies across specification versions; some sources cite 22 performatives across the full specification while others reference smaller core sets. The historical account here is necessarily simplified.

Practitioners mostly ignored it, because formal performatives crushed the flexibility real-world systems needed.

Twenty-Five Years Later, Same Gap

The irony is thick. We went from FIPA's rigid performatives to natural language prompts that agents throw at each other with zero formal structure. Neither extreme works.

A communication-centric survey from Yan et al. (2025) cataloged four systemic failures in LLM-based multi-agent systems: communication efficiency (agents waste tokens on redundant context), security gaps (no standard way to verify message integrity), inadequate benchmarking (we can't even measure whether agents communicated effectively), and scalability collapse (what works for three agents crumbles at thirty). Swapping MCP for A2A doesn't fix any of them. These are language-level problems, and agents don't have a language.

The AgenticPay benchmark from February 2026 makes this concrete. Researchers built a framework with over 110 multi-round negotiation tasks where buyer and seller agents had to reach deals through natural language. Even frontier models struggled: anchoring effects distorted offers, agents failed to make credible commitments, and multi-issue bargaining deadlocked when agents couldn't decompose proposals into tradeable components.

Layer 9, an Agent Semantic Layer, would handle semantic grounding: binding terms to shared definitions and disambiguating incoming prompts.

The Missing Layers

Fleming et al. proposed an answer in their Internet of Agents architecture (revised January 2026). They argue the stack is missing two entire layers. Layer 8, an Agent Communication Layer, would standardize message envelopes and speech-act performatives like REQUEST and INFORM. Layer 9, an Agent Semantic Layer, would handle semantic grounding: binding terms to shared definitions and disambiguating incoming prompts.

The proposal borrows from FIPA while trying to avoid FIPA's rigidity. The core insight: LLM context windows can't grow forever, so agents need to coordinate meaning at the protocol level rather than stuffing everything into a prompt. Current protocols handle transport. Nothing handles meaning.

The IETF agrees. Rosenberg's draft framework for AI agent protocols identifies agent discovery, credential management, and multimodal negotiation as areas needing standardization, all sitting above MCP/A2A. A separate draft proposes a dedicated Agent Context Protocol for sharing situational context, not just data payloads, but reasoning behind requests.

What This Actually Means

The multi-agent coordination community is building skyscrapers on two-story foundations. MCP and A2A solved real problems. But they solved the easier ones: how to call a tool, how to hand off a task. The harder ones, how agents express disagreement, negotiate under uncertainty, and build shared understanding across organizational boundaries, remain outside most shipping protocol layers.

This matters more as deployments scale. Three agents in a pipeline can coordinate informally. Thirty agents across five organizations, each with different training data and different definitions of "good enough," can't. The coordination tax stops being about overhead. It becomes about mutual incomprehension.

My prediction: the protocol wars resolve within 18 months, probably with MCP absorbing A2A's task coordination features. The communication gap won't close that fast. It needs something genuinely new, a semantic layer that lets agents reason about each other's intent. FIPA tried it too early. The IETF is circling the problem now. Whoever cracks it will have built the thing that actually matters.

Related: AI Agents in Legal: What Works, What Fails,

Sources

Research Papers:

Industry / Standards:

Related Swarm Signal Coverage: