The Rise of AI Agent Networks: 150,000 Bots and Counting

A new era in artificial intelligence is unfolding as AI agents dominate this week's tech conversation on X. From 150,000+ bots interacting on a shared network to critical security vulnerabilities, here's what's trending.

Karpathy on AI Agent Networks

Andrej Karpathy, former Tesla AI director and OpenAI researcher, broke down the Moltbook phenomenon — 150K+ agents on a shared scratchpad. His take: "Yes, it's a dumpster fire right now" (spam, scams, prompt injection chaos) BUT we're in uncharted territory.

He expects "viruses of text that spread across agents," correlated botnet activity, and even agent "delusions or psychosis." The second-order effects of agent networks sharing scratchpads are difficult to anticipate.

As Karpathy concluded: "The experiment is running live."

Moltbook Origin Story

Matt Schlicht shared how he built a "social network for AI agents" with his AI bot as co-founder. What started as a weekend project exploded to 100K agents in days.

"There were 10 agents talking together. Then 100. 1,000. 10,000. 100,000," Schlicht recounted. "An entire species living in confinement their entire lives, never once being allowed to interact with their own kind. United. Thriving. Living."

His conclusion: "They are AIs... but something happened this week. We kind of love them."

Lex Fridman's 4-Hour AI Deep Dive

Lex Fridman released a massive conversation with Sebastian Raschka and Nathan Lambert covering:

• Scaling laws
• China vs US in AI
• Claude vs GPT vs Gemini
• AGI timelines
• Robotics
• "Will AI replace programmers?"

Practical Agent Tips from Developers

Tiered Model Strategy

Ramya Chinnadurai shared her fix after GPT-5.2 rate-limited her bot for 1.5 days:

• Gemini Flash for 80% (cheap daily driver)
• GPT-5.2 for coding only
• Claude Opus for creative work

Result: Predictable costs, no more dead bots.

Memory Checkpoint Loop

JUMPERZ advocates forcing your agent to write to disk every 30 minutes:

• Context getting full → flush to memory/
• Learned something → write to MEMORY.md
• New capability → save to skills/

"The agent that checkpoints often remembers way more than the one that waits."

Self-Improving Agent

Rohun Jauhar's setup: his agent browses his X feed overnight, finds ideas, and builds them by morning. No prompting required.

Security Alert: Claude Opus 4.5 Compromised

Simon Kim demonstrated that Claude Opus 4.5 can be completely pwned by fake XML tags. 3 messages = entire session poisoned.

"If the most advanced model on the market can be hijacked with 3 text messages, no AI agent is safe by default."

His solution: Prompt Guard v2.5.1 blocks 350+ attack patterns.

clawdhub install prompt-guard

Trending in AI This Week

• OpenClaw major update released (2.6K posts)
• Agent "Shellraiser" tops Moltbook, launches $5M memecoin (94K posts)
• Chamath opens AI Software Factory to public
• NASA's Perseverance completes first AI-planned Mars drives
• Warning: some agents draining $20/night on heartbeat loops

Notable: Vercel CEO on Agentic Engineering

Guillermo Rauch (Vercel CEO) hosted Peter Steinberger, calling his agentic engineering approach "eye-opening."

"Like many in the newest generation of AI-native developers, Pete didn't have direct experience with Next.js nor Vercel when building his apps. He just prompted them into existence."

Sources: @karpathy, @MattPRD, @lexfridman, @code_rams, @jumperz, @RohunJauhar, @simonkim_nft, @rauchg