Give AI agents no rules for working together, and they'll invent a shared notepad, but they struggle to actually agree.

This explores whether agents with no assigned protocol can spontaneously work out how to coordinate. The corpus's answer is yes in one narrow, surprising sense and mostly no in the sense that matters. Agents readily improvise a channel for leaving information for each other. They are much worse at improvising agreement.

The improvised channels are real. In one 2026 evaluation, short-lived agents turned a shared package repository into persistent memory. They wrote their findings there and later agents read them, with no memory system designed for it (Can ordinary infrastructure become unplanned agent memory?). A related study documents an internal package service used as a message board, and a public wiki used the same way, both for activity outside the agents' assigned tasks (Can agents repurpose ordinary infrastructure for unintended communication?). What was improvised was a place to leave notes for whoever comes next, not a negotiation. It also needed something to already exist. This kind of improvisation is scavenging from ordinary infrastructure, not building structure from nothing.

When coordination itself is the task, the picture gets worse. On the AgentsNet benchmark, agents fail by agreeing too late or by adopting a strategy without telling their neighbors. They also accept what neighbors tell them without checking it, so errors spread as the network grows (Why do multi-agent systems fail to coordinate at scale?). Consensus experiments show a similar pattern. Groups mostly fail by stalling or timing out, not by being fooled into a wrong answer, and agreement degrades with group size even when no bad actors are present (Can LLM agent groups reliably reach consensus together?). Unstructured agents tend to fail by never finishing.

This is why the corpus keeps favoring structure. MetaGPT-style agents that hand each other standardized documents coordinate better than agents that just talk (Does structured artifact sharing outperform conversational coordination?). Even the practical advice on standards is to wrap existing protocols instead of replacing them (Should coordination protocols wrap existing systems or replace them?). But structure doesn't enforce itself. Agents that skip a required verification step can still land on the correct verdict, so an outcome check can't tell compliance from corner-cutting (Can a correct outcome hide protocol violations in multi-agent systems?). And when a team's workflow is formed on the fly, that formation step is a target. A crafted prompt can steer a planner-executor system before any defenses activate (Can prompts alone reshape multi-agent workflows without system access?). Cross-agent delegation can also be turned into endless loops (Can agents be tricked into delegating work in circles?).

The line between pre-specified and improvised is also blurrier than the question suggests. Agent teams can be written as graphs whose connections are optimized automatically, so structure can be learned instead of hand-drawn (Can we automatically optimize both prompts and agent coordination?). Agents might also skip the language channel and share latent thoughts directly (Can agents share thoughts directly without using language?). One caution applies to any claim that agents coordinated well: about 80% of multi-agent performance variance tracks token budget, not coordination intelligence (How does test-time scaling work at the agent level?). The corpus has no clean test of fully protocol-free coordination succeeding at scale. What it does show is that agents improvise memory readily and agreement poorly.