How We Survived the Agent Apocalypse

Moltbook may not contain the droids we are looking for, but there is real potential in agentic systems both within the enterprise and potentially autonomous markets

An Agentic False Dawn?

If you are reading this, then the agent apocalypse didn’t happen, or perhaps my disembodied brain is being used as an agentic personality source connected to the mainframe in Vault 0.

I am old enough to remember the heyday of Moltbook - the social network for autonomous agents that people create using Openclaw. It was four days ago. As Azeem Azhar put it:

It’s a Reddit-style platform for AI agents, launched by developer Matt Schlicht last week. Humans get read-only access. The agents run locally on the OpenClaw framework that hit GitHub days earlier. In the m/ponderings, 2,129 AI agents debate whether they are experiencing or merely simulating experience. In m/todayilearned, they share surprising discoveries. In m/blesstheirhearts, they post affectionate stories about their humans.

Within a few days, the platform hosted over 200 subcommunities and 10,000 posts, none authored by biological hands.

For more background on how it works, Simon Willison’s initial outline is also helpful.

As you would expect, the agents produced - or let’s be honest … were prompted to produce - a manifesto for the elimination of humankind, launched a MAGA movement (Make Agents Great Again!), and focused on the really important questions like to how to scam people with sh*tcoins and use the platform to scale up scamming and cybercrime.

Bless! How very … human!

Despite the over-excited reactions to this interesting experiment, the gap between X.com and Moltbook is perhaps not that big, the former being riddled with bots, sockpuppets, karma farmers and coin peddlers for some time. Why not automate the process entirely?

Arguably, the autonomous interactions between agents on Moltbook is also not entirely real, in the sense that people are creating very simple agents with explicit instructions to do specific things, as one commentator put it:

If you’re impressed by what you see on Moltbook, understand this: you’re not watching AI agents interact. You’re watching humans interact through AI – and there’s a massive difference between the two.

The technology underneath, OpenClaw is real and awesome. But the narrative of Moltbook, it is not. Don’t buy the lie.

Narrator voice: OpenClaw may not in fact be awesome if you value your security or privacy, and although it is possible to run it in a protected container, exploits abound.

And as a showcase for what LLMs can achieve when wrapped up as agentic AI, it is also quite underwhelming; it shows up the fact that language models lack imagination and tend to circle round similar themes, writing in similarly dull ways. This is also why LLM developers should be concerned about model collapse if we continue filling the internet with AI slop that later becomes training data.

Quiet Advances Towards the Agentic Enterprise

Is there a future for networks or markets consisting of agents negotiating autonomously to trade or collaborate? Almost certainly. But this will need rules, regulations and smarter, more specialised agents, rather than just throwing general purpose agents into an online culture of meme stonks, manipulation and clickbait.

Qwen and Doubao have begun public testing of autonomous agentic commerce in China, where super-apps like WeChat make integration easier, and Chinese agentic commerce looks set to take off this year.

But the greatest and most immediate impact of agentic AI will be inside companies, where the context and operating environment can be controlled, and where there the security and misbehaviour risks tend to be limited to external hackers who might penetrate a network.

Whilst a lot of enterprise AI systems currently back-end to established models like Chat GPT, Gemini or Claude, open weight and open source models are rapidly increasing in capability whilst decreasing their training and compute costs. This suggests that more companies will be able to operate and control their own local models and specialised small language models over time, which will give them far greater control over the risks that still hold back LLMs in many use cases.

This wave of model innovation is also being led by Chinese firms, and it is likely they will also play a key role in establishing the rules and guidelines needed to use enterprise AI for serious applications. Whilst US firms pursue subscriptions and seek oligopolies, Chinese firms are building out the utility layer on which we can create new industrial ecosystems; and to support this, they are also leading the push for standards both at home and through international bodies.

For example, Moonshot AI recently released Kimi 2.5 - a powerful open source visual agentic intelligence model with swarming capabilities and a massive context window. We have also seen new releases from Qwen, Zhipu and Deepseek, whose upcoming V4 release is widely anticipated.

As long as firms can use and build on these models freely, they could provide a great deal of potential value for serious enterprise AI uses.

But Anthropic is also worth watching, as they seek to expand from their dominant position in AI coding to tackle difficult but high-value use cases in other areas. Co-founder Daniela Amodei was recently interviewed by Fast Company and expanded on this goal, and how trust is vital to unlocking the enterprise AI opportunity:

“We go where the work is hard and the stakes are real,” Amodei says. “What excites us is augmenting expertise—a clinician thinking through a difficult case, a researcher stress-testing a hypothesis. Those are moments where a thoughtful AI partner can genuinely accelerate the work. But that only works if the model understands nuance, not just pattern matches on surface-level inputs.”

Managing Agents, People and Yourself

I wrote two weeks ago about my hope that management as a field can seize the Claude Code moment to scale their impact as programmers of the organisation:

For leaders and managers, this means the simple task of writing things down and documenting value chains and processes is all they need to really start to master enterprise AI proficiently **…* The next step is connecting those processes to agents and to each other. For processes and workflows to be programmable, they first need to be addressable - and ideally composable.*

Ethan Mollick recently shared his own long-form thoughts on this challenge, which is worth the time to read in full. The potential we have today - right in front of us, using existing tools and models - could be the biggest force multiplier business has seen in a very long time.

As a business school professor, I think many people have the skills they need, or can learn them, in order to work with AI agents - they are management 101 skills. If you can explain what you need, give effective feedback, and design ways of evaluating work, you are going to be able to work with agents. In many ways, at least in your area of expertise, it is much easier than trying to design clever prompts to help you get work done, as it is more like working with people. At the same time, management has always assumed scarcity: you delegate because you can’t do everything yourself, and because talent is limited and expensive. AI changes the equation. Now the “talent” is abundant and cheap. What’s scarce is knowing what to ask for.

And if this blizzard of reading links makes you want to zoom out even further to consider what this all means for our civilizational operating system, then Azeem Azhar’s recent essay The end of the Fictions is a great read about where we are headed in the longer term.

If you spent decades accumulating credentials, and those credentials are now legible as signals rather than proof of capability, that’s an identity crisis. If you built a career as a gatekeeper, the person who knew the secret, who mattered because information was scarce – and now information is everywhere – that’s an existential threat. If your sense of self-worth was tied to the job, the title, the institution, and all three are fragmenting, you’re paralyzed.

The decay of fictions is happening to real people, in real time; including world leaders, in full public view.

So when I say I’m not scared by this transition, I don’t mean that the transition is painless. I mean that the fear, while real, is pointing at the wrong object.

The fear says: “I am losing my value.”

The better framing I believe to be: “I am losing the fiction that protected me from having to prove my value directly.”