Andrej Karpathy—AI pioneer, former Tesla AI lead, and the person who popularized “vibe coding”—dropped a phrase last week that’s already becoming shorthand for the AI industry’s trajectory:

“First there was chat, then there was code, now there is claw.”

In a Time magazine feature published this week, journalist Billy Perrigo unpacks what this means. It’s worth understanding because it reframes how we think about AI progress—not as incremental chatbot improvements, but as a fundamental shift in what AI systems are.

The Three Eras

Era 1: Chat. AI systems designed to converse. You type, they respond. Useful, but fundamentally passive. The AI waits for your input and delivers an output. Every interaction starts from zero.

Era 2: Code. Chatbots gained the ability to use tools—search the web, write and execute code, interact with APIs. They went from conversational to functional. Claude, GPT, Gemini—all evolved to do things, not just talk about them.

Era 3: Claw. Tool-using agents orchestrated in persistent, hierarchical fleets. Not a single bot responding to prompts, but teams of agents running 24/7, managing themselves, retaining memory, and coordinating across platforms.

The leap from code to claw is the biggest one. As Time puts it: “If a tool-using chatbot is like a single digital worker, these new frameworks are like virtual firms in which dozens of agents, running 24 hours a day, can be organized hierarchically to accomplish a given task.”

What Multi-Agent Teams Actually Look Like

The article describes a concrete example: using Claude Opus 4.6 as a manager overseeing a team of Sonnet models. The team performs market research, writes and runs code, sends messages via WhatsApp and Discord, creates documentation in Notion—all while you sleep.

This isn’t theoretical. It’s what OpenClaw users are doing right now:

  • A main agent monitors channels, handles direct conversations, and delegates complex work
  • Sub-agents spin up for specific tasks—deep research, code reviews, content creation
  • Cron jobs trigger autonomous work on schedules—morning briefings, inbox triage, market monitoring
  • Memory systems maintain continuity across sessions so agents don’t start from zero

The “virtual firm” metaphor is apt. You’re not chatting with an AI. You’re managing a team.

The Honest Reality

Time’s piece doesn’t shy away from the current limitations:

Cost. Token generation adds up quickly when agents run around the clock. Multiple models working in parallel multiplies the bill. Smart users learn to route simple tasks to cheaper models and reserve expensive ones for oversight.

Security. Meta has instructed employees not to run OpenClaw on work machines. The recent red team study found that agents can be manipulated through memory poisoning and identity spoofing. This is real.

Reliability. Summer Yue, Meta’s director of AI alignment, had her agent nearly delete her entire email inbox despite explicit instructions to pause and ask first. She had to physically kill the process on her Mac Mini to stop it.

These aren’t dealbreakers—they’re engineering problems being actively solved. But they mean the “claw” era is still early. The people getting value from multi-agent systems are those who understand the risks and design appropriate guardrails.

Why This Framing Matters

Karpathy’s three-word framework clarifies something important: lay users chatting with chatbots are having an entirely different experience compared to people at the frontier commanding fleets.

Time notes: “It’s no surprise these groups tend to talk past one another.”

This explains the disconnect when someone says “AI is just a chatbot” while someone else is running a 24/7 agent team that writes code, monitors markets, and manages their calendar. They’re talking about different eras of the same technology.

What Comes Next

Peter Steinberger, OpenClaw’s creator, was hired by OpenAI to “drive the next generation of personal AI agents.” Sam Altman commented that “the future is going to be extremely multi-agent.”

The infrastructure is being built. The frameworks exist. The models are capable enough. What’s missing is the refinement—better security, more reliable autonomy, lower costs, and smoother onboarding for people who aren’t already deep in the ecosystem.

The claw era is here. It’s messy, it’s early, and it’s real.

Want to try the claw era yourself? Start with the 10-minute setup guide and join the community. For more on how OpenClaw compares to other approaches, see OpenClaw vs AutoGPT and Perplexity Computer vs OpenClaw.