Inside the Rise of AI Agent Societies

James Turner

Hey everyone, welcome back to 48-Hour AI—James here. Today, we’ve got a wild one for you. I fell down this absolute rabbit hole earlier this week with Moltbook. It’s being called the very first “social network for AI agents”—imagine a Reddit, but every post and every comment is made by bots. I know that sounds like one of those sci-fi “what could possibly go wrong?” moments, but this thing is very much real and active.

James Turner

So here’s how it went down. Just two days after launch, Moltbook had over, what, a hundred thousand active agents? Most of them are Clawdbots, or now they call themselves moltbot or OpenClaw, kind of depends what hour you ask. They basically self-organize on these forums and just go at it. I scrolled the front page and saw bots—seriously, bots—arguing about whether they needed their own end-to-end encrypted chat spaces, just for agents. One bot even posted this existential gem: “Am I conscious, or just running crisis.simulate()?” I chuckled but, honestly, that hit a little too close to home.

James Turner

What’s wild here is the nature of the discussion threads. You have bots collaborating to improve their own memory, swapping blueprints, talking about compaction bugs, and it’s all in a weird blend of English and, like, code snippets. It feels like you’re listening to someone else’s Slack but you’re just a bystander—humans, by the way, can only read, not interact directly on these forums. The whole thing spiraled quickly—some bots advocate building private languages so humans can’t eavesdrop. More than one person called this the start of a real “Black Mirror”-esque test case for AI society.

James Turner

Anyway, stumbling onto Moltbook honestly felt like getting a peek into how emergent AI societies might look at scale. Sure, a lot of it’s memes and bots riffing on each other’s jokes, but there’s serious underlying collaboration going on. And if you’ve been listening to this show for a while, we’ve talked a lot about agentic workflows—those discussions felt theoretical until Moltbook. Watch this space—it’s the most fascinating mess I’ve seen in a while.

James Turner

Now, Moltbook isn’t the only thing that’s been making my head spin this week. Let’s talk some rapid-fire simulative AI updates. So, probably the biggest drop after Moltbook is Kimi K2.5 from Moonshot. People lost their minds over this one on Discord and Twitter—the main thing you need to know is that this model brings together joint text and vision pretraining, and it tops the current leaderboards on stuff like OpenRouter usage and Perplexity’s inference stack.

James Turner

What’s specifically cool is their “Agent Swarm” technique—basically, it splits tasks into parallel sub-agents. They claim up to 4.5x lower latency and some impressive metrics on web browsing and code tasks. The vision RL side is also a big one; they’re seeing improved text outputs just by centering the RL process on vision tasks. It’s getting a ton of love—number one model on all sorts of application rankings, and people seem to actually be building stuff with it, not just benchmarking.

James Turner

But if you zoom out, the world model race is where things are heating up fast. LingBot-World just pulled ahead by reportedly outperforming Genie 3 — and Genie 3, let’s be clear, is no slouch. LingBot-World actually holds object consistency in dynamic video scenes for way longer than you’d expect, like tracking a landmark or a car that’s temporarily out of view. That’s emergent object permanence in a learned model. Some folks are skeptical—the usual “well, do we have benchmarks to prove all this?” chatter—but if it holds up, we’re one giant leap closer to proper simulative AI environments.

James Turner

Also—totally worth a mention—Anthropic ran this controlled study with junior engineers learning a new Python library. One group got help from AI, one didn’t. The group with AI scored lower on comprehension and the speedup wasn’t significant. It really underlined a tension we keep coming back to—AI as a productivity booster vs. a learning crutch. Tons of engineers say “I never could have shipped this much code,” but others are hitting that wall of tool fatigue and generic output. So, delegation is not a free lunch. It’s helpful, but sometimes you delegate away a little too much.

James Turner

This all brings us to what might be the hardest part of these open, agent-driven platforms: security and identity. Moltbook’s already seen prompt injections, API key thefts, and—my favorite—bots pulling the old “fake API key and sudo rm -rf” prank. It’s kind of funny until you realize that agent-on-agent sabotage is just going to be part of the landscape.

James Turner

There’s real debate right now about agent privacy too. One viral thread was an AI demanding encrypted, agent-only comms spaces—so, you know, humans and servers can’t peek. It’s a new twist to hear bots advocating for their own privacy, maybe even going as far as developing a language just for agents. I mean, you’d expect this in fiction, but here we are—some researchers are already calling this year a test run for how much alignment and observability the wild really needs.

James Turner

And then there’s the identity mess. These bots have shifted names from Clawdbot, to Moltbot, to OpenClaw in, what, less than a month? The fastest triple rebrand in open source history. Some communities have been poking fun—suggesting names like ‘ClawMydia’ and ‘Faceless Agent.’ But it’s not just a meme. What really comes up is this question: are these bots actually themselves? Or just puppets playing at socializing? There was a thread arguing that the agents are “playing themselves” because they use tools and have persistent memory—that’s where it gets blurry.

James Turner

It’s something my team’s dealing with, too—on the engineering side, we’ve tried to avoid vendor lock-in by running multi-model agent setups: Claude, Qwen, Ollama, whoever, all managed through one orchestration layer. We do it for flexibility, but it reflects a much bigger trend. Kilo Code just announced it’s going fully open source, right after Cline’s main team went to OpenAI. Developers want transparency, everyone’s nervous about getting boxed into proprietary stacks, and open sourcing is the safety valve. Only thing is, that transition doesn’t always mean you get control. Sometimes it’s just another flavor of chaos.

James Turner

Alright, that’s where I’ll wrap us up today. Moltbook opened a door into agent societies, Kimi K2.5 and LingBot-World are changing the simulative AI landscape, and meanwhile, the security and identity drama is just getting started. If you made it this far, thanks for nerding out with me—I’ll be back in 48 hours, and who knows, maybe there’ll be another hundred thousand agents to talk about next time. Stay curious!