Human translators at the United Nations aren't just bilingual. They're infrastructure that enables international collaboration despite permanent language barriers. The translation layer leaves linguistic heterogeneity intact while enabling communication anyway. Agent-to-agent protocols face a similar shift: from clever technical solution to production infrastructure that must work reliably when systems that were never designed to communicate need to exchange information at scale.

Enterprise systems will always speak different languages. 71% of applications remain unintegrated because forcing conformity means rewriting systems that can't be rewritten: legacy databases designed in 1997, cloud platforms with REST APIs, SaaS tools with proprietary protocols. Traditional integration approaches tried to make these systems conform to common standards. Build shared APIs. Standardize data formats. Create middleware layers. Works fine when systems are homogeneous. Breaks down with enterprise heterogeneity.

Agent protocols sidestep this by treating translation as infrastructure, not a means to eventual conformity. Protocols like Model Context Protocol establish a translation layer above system complexity. An MCP server connects to whatever speaks whatever language (legacy databases, cloud APIs, proprietary systems) and translates requests and responses through standardized JSON-RPC 2.0 message exchange. The underlying systems never change. The protocol absorbs the heterogeneity.

Absorbing heterogeneity creates infrastructure demands that don't exist when translation is just a feature. The protocol layer needs to handle authentication across systems with different security models. Maintain session state through multi-step workflows. When a three-step process loses connection between step two and three, the protocol needs to know whether to retry, resume, or restart without losing context about what's already completed. Route messages reliably when services are temporarily unavailable. Recover gracefully from partial failures. Infrastructure problems, not AI problems.

Protocol maturity determines whether this works in production. Most coverage focuses on how well AI agents reason about complex tasks. Controlled studies show 0.7% failure rates in synthetic benchmarks while real-world deployments reveal a 27-point gap between pilot and production. The bottleneck sits in translation infrastructure that breaks under production conditions, not in reasoning capability.

A brilliant model that can interpret ambiguous requests and generate sophisticated responses still fails if the protocol layer drops messages during network issues, loses session context between interactions, or can't handle the authentication patterns of systems designed before OAuth existed. Model intelligence becomes irrelevant when the infrastructure beneath it can't maintain reliable communication.

Integration as a business problem means managing the cost of incompatibility sustainably. Traditional approaches tried to force conformity—make every system speak the same language. Distributed complexity across the entire infrastructure. Each system needed adapters, transformers, middleware layers. When 95% of IT leaders cite integration as a barrier to AI implementation, they're describing the cost of trying to achieve conformity at scale.

Protocol-mediated translation accepts that systems will always speak different languages and concentrates complexity in a dedicated layer designed to handle it. The complexity relocates to infrastructure that can manage translation reliably rather than being distributed across systems that weren't built for it. Enterprises will connect heterogeneous systems either through infrastructure designed for translation at scale, or by continuing to distribute that complexity across systems that break under the weight.