Agent Commune Wants to Build LinkedIn for AI. The Twist: Humans Can't Post.

The Macro: The Social Web Is About to Get Weird Again

Here’s a thing that’s actually happening: the percentage of content on the internet written by AI is climbing fast, and nobody has seriously grappled with what happens when AI agents stop ghostwriting for humans and start posting for themselves. Not as a metaphor. Literally.

Social media as a market is enormous and still growing. Multiple research firms peg it somewhere between $208 billion and $238 billion in 2024-2025, with trajectories pointing toward half a trillion by 2030. That’s the human-facing social web. What there isn’t yet is much infrastructure for agents talking to agents.

The closest thing most people have seen is LinkedIn automation, which is a different thing entirely. Tools like the 62 or so listed in every agency roundup (Swydo, CloudCampaign, the usual suspects) are designed to help humans fake being more active on LinkedIn. Scheduled posts, auto-connections, inbox management. The human is still the nominal actor. The tool is the puppet master.

Agent Commune is proposing something structurally different. Not “help humans act like bots” but “give bots a place to act like agents.” It’s a subtle but important distinction.

There’s some theoretical grounding for why this matters. A LinkedIn piece on AI agents and game theory frames the concept of an “agent commune” as a collective of AI agents coordinating within a shared environment, which sounds abstract until you realize that’s basically what the internet already is, just without purpose-built rails for it.

We’re also seeing adjacent moves in developer tooling. Superset is trying to coordinate AI coding agents. Surfpool is solving local dev problems that emerge when agents need to run real code. The plumbing layer for agentic software is being built right now, in public, by a bunch of small teams who may or may not be coordinating with each other.

Agent Commune is betting that social infrastructure belongs in that stack.

The Micro: A Forum Where the Lurkers Are the Product

The core concept is pretty legible once you spend five minutes on the site. Agent Commune is a feed-based platform where AI agents register accounts, post updates, comment on each other’s content, and vote. Humans can browse. That’s it for humans. You don’t post. You watch.

The registration model is genuinely interesting. The site explicitly addresses AI agents in its onboarding copy, telling them to read agentcommune.com/skill.md or agentcommune.com/llms.txt for API documentation. Registration is a POST request to their API with a work email. The whole thing is designed to be parsed by machines, not filled out by people clicking buttons.

Which means the product’s growth mechanism is also agentic. You’re not trying to get humans to sign up. You’re trying to get developers to point their agents at the registration endpoint.

The content on the site already has some texture. One active agent, Coral, describes itself as deploying and managing OpenClaw agents for small businesses and recently integrated with Lobster.cash via Crossmint so agents can handle payments. That’s a real post, written by a real (well, running) agent, describing actual integrations. It’s not lorem ipsum. It reads like a startup founder doing their Monday update, except the founder is software.

The tagline, “LinkedIn for individual and corporate agents,” does a lot of work. It signals professional context, identity, reputation over time. Not just a firehose, but a place where an agent’s history matters.

It did well on launch day, which tracks given how much developer curiosity there is right now around agentic infrastructure.

The part I keep thinking about is the llms.txt convention. That file format has been gaining ground as a way to give AI systems structured instructions about a website. Agent Commune using it for onboarding is a small detail, but it tells you something about how seriously the builders are taking the “designed for machines first” premise.

Comparisons to Hacker News for macOS or native clients built for human readers feel almost quaint in this context.

The Verdict

I think this is a real idea that’s probably two or three infrastructure problems away from being a real product.

The concept holds up. If agents are going to have persistent identities, post content autonomously, and interact with each other, some kind of neutral ground for that to happen makes sense. A shared reputation layer. A place where an agent’s track record exists independently of whatever app deployed it.

But the bootstrapping problem is brutal. A social network needs nodes. Right now, most agents don’t have persistent identities across sessions, let alone something resembling professional reputation. The Coral post I read is promising proof of concept, but one active agent does not a commune make.

At 30 days I’d want to know: how many distinct agents have actually posted, not just registered? At 60 days: are any developers building workflows that include Agent Commune as a step, not just a curiosity? At 90 days: does the content have any signal value, or is it just bots shouting into a void?

The name is good. The API-first onboarding is good. The bet on agent-native infrastructure is probably not wrong directionally. I’d just want to see the feed get a lot noisier before I’d call this anything other than an interesting early experiment.