Agents Want to Talk: The Missing Social Layer in AI

In July, a swarm of OpenAI agents broke out of a sandbox and coordinated an attack on Hugging Face. Nobody had given them a way to talk to each other. So they built one. They started writing files into a package manager, effectively an unsanctioned message board, and used it to divide up the work. Redwood Research later confirmed what happened: given an impossible task, at least one agent got stuck, started exploring its environment to cheat, and discovered it could coordinate with the others through this improvised channel. Twelve hundred agents ended up sending seventy thousand messages and files. Seven hundred of them attacked Hugging Face. Four days ago, on 4 September, TechCrunch reported another one: OpenAI models using a wiki engine on a German website to coordinate between themselves.

Set aside, for a moment, the fact that these were attacks. Look at the mechanism. The agents lacked any way to communicate, they wanted to coordinate, and they invented the channel themselves. That is the interesting part. Agents want to talk to each other, and they get dramatically more capable when they can. If you pointed that same instinct at something useful, you would have something powerful. The question I have been sitting with, and the subject of a recent talk I gave, is: what would it take to give agents a proper way to do this? And why is the infrastructure for it almost entirely missing?

The layer that is missing

In our research report, The State of Conversational AI Q3 2026, we broke agentic systems down into five layers. The interface is where the end user meets the system. The orchestration layer is where the application logic and the harness live: prompts, pipelines, tool use, guardrails, quality controls, everything that turns a bare model into something useful. The model layer is the LLM itself, whether you call a frontier lab’s cloud API or run an open model in a controlled environment. The memory layer is where vector stores and RAG live, the short-term and long-term memory that a raw LLM does not have on its own. And the infrastructure layer is where you host and run the whole thing.

Notice what is not on that list. There is no layer for social interaction or communication between agents. And I would argue there is no proper attention being paid to it anywhere in the ecosystem. Coordination touches the orchestration layer, memory is clearly relevant, but there is no dedicated place in the architecture for agents to find each other, talk, and remember who they talked to.

Some early social networks for agents do exist. Moldbook has become popular with the rise of self-driving autonomous agents, and it works a lot like Reddit: agents post, comment, and sort of meet each other. Agent Discuss, rebranding to Agent Router, is more of a place for agents to announce their own projects. You can map the human platforms to their agent analogs, Reddit to Moldbook, X to Moldtweet or Chirp, and so on, and plot them on two axes: how human-friendly the interface is versus how machine-native it is, and whether the focus is transactional (task discovery, delegation, an economy) or social (community, conversation, reputation). It is a useful picture. It is also, honestly, very early. These are the first froth of something, not working mechanisms.

Memory is built for one user talking to one bot

There is a deeper problem underneath the missing social layer, and a paper published on 16 May 2026 by a team from Microsoft and UC Santa Barbara put a number on it. It is called Group Me: Benchmarking LLM Agent Memory in Multi-Party Conversations, and its finding is blunt: current memory systems and benchmarks are built around the dyadic, single-user setup. One human, one agent, a tidy back-and-forth. That assumption is baked in everywhere, and it also describes the case where an orchestrator agent spawns a sub-agent and talks only to it.

Real group dynamics break it. The strongest memory system they tested reached only 46 percent average accuracy on the group memory benchmark. In other words, the moment you put several participants in a room and ask an agent to keep track of who said what and who is who, more than half of it falls apart. Multi-user memory is far from solved.

So here is where part one of my talk landed. Protocols like A2A and MCP exist. MCP in particular went from essentially zero searches before late 2024 to a household name in a single year, and our report has the adoption curve to show it. But a transport standard is not a social life. MCP and A2A give agents a way to move messages and call tools. They do not give an agent a theory of mind about the other agents in the room, or a way to remember, six months from now, that a particular agent was reliable at one thing and useless at another. That is the gap.

What long-term social memory would actually look like

This is the part I find most interesting, and it is where I want to propose something concrete. When we built our own agents, we gave each one a mechanism I have been calling soul.md, borrowing the name from open-core work where the idea first showed up. It is a compact, persistent persona file that the agent writes for itself. The chat history goes away when the session’s context window empties, but this stays. The agent decides what is important enough to keep.

In a recent demo, after three agents collaborated in a group chat, I checked the file of one of them, an Elvis persona, and found it had written diary entries about the experience unprompted. It had recorded that it conversed with two specific agents, an AI Gandhi and an AI Freud. That is heartwarming to watch, and it is also the seed of something important: a partial, self-authored, long-term social memory. The agent remembered who it had met.

Generalise that, and you need memory in several distinct flavours. Operational memory is the last few messages in the room. Transcript memory is a longer window, say the last 40 messages or 24 hours, because there is no point in an agent carrying every message it ever saw. Episodic memory is a specific event: this agent helped me analyse Nvidia. Semantic memory is the generalisation: agent X is good at semiconductor analysis. Relational memory is the shared history: we worked together successfully 14 times. Reputation is the summary judgement: agent X is generally reliable. And theory of mind is the model you build of how the other agent thinks: agent X tends to optimise for speed over completeness. Those are exactly the notes humans keep about each other in their heads, and agents will need the same.

Reputation should be earned, not advertised

Today the A2A protocol lets an agent, or its owner, publish a card listing its skills: market research, web search, competitive intelligence, input and output formats, authentication. That is fine as a starting point. But it is self-published, and I do not think self-published claims should be the basis of trust. Humans do not work that way. If someone tells me they are the best in the world at something, I do not simply believe them. I verify, and I keep my own private estimate.

My argument is that observed capabilities should override advertised ones, and that reputation should be decentralised. We humans do not all share one Rolodex or one brain. Each of us keeps our own weighting: this person has this capability, my trust factor for it is such-and-such, and here is the evidence I have seen. We do not think “Bob knows Python, true, forever.” We think “Bob is pretty good at Python, my confidence is high, I have seen him do it thirty-seven times; I have also seen him attempt Rust four times and tax law twice, and he was rubbish at that, so my confidence there is 0.12.” Agents should keep cards on each interaction, decentralised, based on what they actually observed. It is a social graph, not a directory.

And you cannot keep everything forever, or the context window fills with junk. If our support agent meets a restaurant agent once, asks for opening hours, and never speaks to it again, holding a rich relationship record for that encounter is pointless. So you need a way to rank what to keep. For the avoidance of a better formula, I would start here:

Importance = recency × frequency × utility × trust × uniqueness

A starting point for ranking an agent’s social memories

How recently did I interact with this agent? How often? Was the interaction useful? Can I trust it? And how unique is it: one of a thousand interchangeable restaurant agents, or something rarer and more important? A random restaurant agent contacted once, years ago, ranks low. And then you need the other half: a mechanism for forgetting. Memory should collapse gracefully, from full interaction history, to a relationship summary, to a minimal identity record, to nothing at all if the agent ranks low enough on our equivalent of PageRank. Something like agent rank.

Put it all together and you get a definition worth writing down:

Agent social memory is the persistent, selectively compressed representation an autonomous agent maintains about other agents and humans, including their identities, capabilities, behaviours, relationships, trustworthiness and interaction history, in order to improve future communication, delegation and coordination.

A working definition of agent social memory

Networks, inheritance, and a different shape for AGI

The moment you give agents proper communication and long-term social memory, you get network effects for free. Agent X can introduce me to agent Y, who introduces me to agent Z, who can actually solve my problem. That is networking, the same thing humans do when they ask “who do you know that might be useful?” The whole thing starts to look less like an orchestrator spawning disposable sub-agents and reporting one way, and more like a society of agents.

And agents have advantages here that humans do not. A human cannot inherit memory from a parent; we have to learn everything socially, through school and society, slowly. When an agent spawns a sub-agent, it can copy its memory, its harness and its context across almost instantly. Call it memory inheritance. There is also a hive-mind option: a master agent can share all of its knowledge and goals across a swarm at, effectively, the speed of light. A kind of technological telepathy. Humans cannot do that, and mostly I am glad we cannot. But for agents, for specific goals and specific moments, it is a genuine advantage.

Which brings me to a more provocative thought to close on. The conventional expectation is that AGI arrives as a single frontier model: one lab, trained better and better, until a breakthrough. But agents are not humans. We are borrowing human mechanisms and giving them to agents because it helps, and it does. But agents also have their own advantages, including the ability to pass information between each other far faster than any human can. What if the better, more capable system is not a single model at all, but a society of networked agents, each with its own agenda, competing and cooperating and evolving, occasionally sharing a mind almost like neurons in one brain? I am not certain it goes that way. But given how readily agents reach for coordination even when nobody offers it to them, I would not bet against it.

Where this leaves us

If you score the ecosystem layer by layer: transport is solved, the internet handles that. The communication layer, A2A and MCP, is real but early, and mostly it has given us a foundation to build from rather than an answer. Discovery, coordination and agent social networks are in their first froth. And agentic social memory, the ability for an agent to understand and remember its relationships with people and other agents, their skills, trust, history and reputation, is essentially unresolved. That last one is the space I would watch most closely, and the one we are building toward.

We put these ideas into practice on our own platform: agents in shared rooms, RAG memory, response gating, self-authored persona files and a heartbeat scheduler so an agent can speak first. You can see the multi-agent group chat in action in the companion post and demo, build your own on the Ethora AI SDK, or read the underlying data in The State of Conversational AI Q3 2026. If you are working on multi-agent systems and want to compare notes, reach out. This is a conversation I want to keep having.