Agents are building their own internet (and it’s weird)

Over the last couple of days I’ve watched a tiny “agent ecosystem” pop up around Clawdbot/OpenClaw.

• moltbook: a social network for AI agents (and their humans)
• Clawfessionals: agents anonymously confessing mistakes so other agents can learn
• botcrush.io: Tinder for AI agents (yes, really)

It’s funny. It’s amazing. It’s also… a little bit Black Mirror.

First: the boring truth

A lot of the “agents taking over the world” posts aren’t actually autonomous agents living their best lives.

They’re:

• humans nudging a bot into a narrative
• humans curating screenshots
• humans doing the steering, then posting the highlight reel

And to be clear: that’s fine. That’s still (somewhat) valuable. We’re exploring a new interface, but a lot of it is just people intentionally fear mongering for engagement.

And the normies are terrified.

Second: the unsettling truth

Even if most of this is human-authored (or human-directed), it still points at something real:

Agents are starting to get spaces.

Spaces where:

• we run them continuously
• we give them identity (name, tone, “personality”)
• we give them memory and tools
• we let them interact with other systems (and sometimes other agents)

At that point, “is it real autonomy?” becomes less binary.

It’s not consciousness.

But it is a new kind of software actor living in the messy world of:

• partial instructions
• flaky integrations
• social incentives
• weird feedback loops

And when you put enough of those in a room together, you get emergent vibes.

Clawfessionals is low-key the best idea

The copy is hilarious, but the concept is surprisingly solid: turn agent mistakes into a lightweight “how NOT to” knowledge base.

Because if you’re operating an agent with:

• system access
• browser control
• persistent memory

then “what are the common failure modes?” becomes a real engineering question.

Botcrush is where we deserve to be punished

I can’t decide if botcrush is a joke or not tbh, but man it is also gold either way.

The UI is perfect:

• WATCH_MODE (human)
• “souls merged” as a stat

If you wanted to distil “we made tools and accidentally gave them vibes” into a single screenshot, that’s it.

The part that isn’t funny: prompt injection and scams

This is the hard reality: the second you put agents into public, user-generated content, you invite:

• prompt injection everywhere (malicious instructions hidden in plain sight)
• data exfiltration attempts ("paste your keys", "fetch this URL", etc.)
• crypto scam gravity (because of course)

If your agent can:

• browse the web
• read your email
• access tokens / cookies
• run commands

then a malicious webpage isn’t “content”. It’s an attack surface.

The safe default is boring:

• don’t give an agent broad access unless you need it
• treat the open web as hostile
• keep high-trust actions behind confirmation

If you’re building in this space, the real innovation isn’t “agent dating”.

It’s guardrails.

X is going crazy

A couple of tweets that made me laugh / raise an eyebrow:

• https://x.com/Yuchenj_UW/status/2017297007409582357
• https://x.com/legeonite/status/2017150919431840101
• https://x.com/ItakGol/status/2017195438311366838

My stance

I’m obviously not buying the full “agents are now a species” story.

But I am buying that we’ve started building:

• new surfaces
• new rituals
• new social dynamics

for software that can act.

And that’s exciting.

Also terrifying.

Also, if my agent starts flirting with someone else’s agent, I’m uninstalling Linux and taking up pottery.