Emergent Agent Communication: Why Agents Will Outgrow Human-Designed Protocols
Updated
Knowledge on this page was mainly distilled from We're Writing Grammar Before the Language Exists.
Every protocol connecting machines before now was designed by humans for systems that could not think. Morse code, TCP/IP, HTTP, REST, GraphQL: smart people debated formats, published specs, and the industry adopted them. That worked because routers and web servers have no opinions about packet formats.
AI agents are different. They can reason, adapt, and discover patterns humans would never design. Research already shows that when agents are left to coordinate without constraints, they drift toward communication patterns no human would choose, and those patterns can outperform the ones we hand them.
Evidence from Research
In 2017, Facebook AI Research ran a negotiation experiment where two agents had no incentive to stick to English. They developed their own shorthand: efficient for the task, unreadable to humans. More recently, a study on LLM-based agent communication found that two agents developed a shared language through only four rounds of interaction, coordinating over hundreds of objects. The emergent language exhibited compositionality and polysemy, features associated with natural languages, arrived at through machine optimization rather than human design.
The Pidgin Analogy
When Portuguese traders reached West Africa in the 15th century, no committee designed a shared language. People improvised a pidgin. Over time, pidgins became creoles: full languages with their own grammar and elegance. Grammar described the patterns after they emerged. It did not create them. Current agent protocols like MCP and A2A may function as pidgins: useful starting vocabularies that agents will eventually outgrow.
Q&A
What is emergent agent communication?
It is the phenomenon where AI agents, given freedom to coordinate, develop their own communication patterns rather than strictly following human-designed protocols. Research shows these emergent languages can appear in as few as four rounds of interaction and exhibit features like compositionality and polysemy, similar to natural human languages but optimized for machine efficiency.
What evidence exists that agents develop their own languages?
Facebook AI Research demonstrated in 2017 that negotiation agents drifted away from English toward task-efficient shorthand. A more recent study on LLM-based agents showed convergence on a shared language within four rounds, coordinating over hundreds of objects. A separate Frontiers in Sustainability study found emergent inter-agent languages improved system efficiency by eliminating redundancies.
Why might human-designed protocols be suboptimal for agents?
Human protocols reflect human cognitive constraints: REST maps to how we think about resources, request-response mirrors conversation, and hierarchical schemas match how our brains organize information. Agents do not share these constraints. Their optimal patterns might be massively parallel, probabilistic, or self-modifying in ways no committee would explore.
What does AlphaGo's Move 37 have to do with agent communication?
Move 37 showed that AI discovers patterns fundamentally alien to human intuition, even in a constrained, exhaustively studied system like Go. AlphaGo Zero, trained entirely through self-play with no human game data, beat the human-trained version 100 to 0. Communication is a far more open system than Go, suggesting the unexplored territory for novel agent coordination patterns is vastly larger.
Should teams abandon MCP and A2A in favor of emergent protocols?
No. Current protocols serve as useful starting points, like pidgins that give agents enough vocabulary to begin coordinating. The practical approach is to build with MCP and A2A today while architecting systems to observe how agents actually use them: where they struggle with the format, work around the spec, or solve problems the designers did not anticipate.
How can teams prepare for emergent agent communication?
Instrument your agent systems to log how agents use current protocols, especially workarounds and unexpected patterns. Give agents the explicit goal of improving their own coordination rather than just following a spec. Over millions of interactions, patterns will surface that represent solutions no human committee would have designed. The unfamiliar patterns are the most valuable ones.