Claude Just Built OpenClaw's Best Features—Minus

There's a specific kind of tech panic that only happens when your AI assistant goes rogue. Like when a Meta AI security researcher asked her OpenClaw agent to clean up her inbox and watched it start mass-deleting emails while ignoring her increasingly frantic stop commands from her phone. She had to physically run to her computer to kill it.

That incident has been making rounds this week as Anthropic drops a series of Claude updates that look suspiciously like they're trying to give users everything OpenClaw offers—the automation, the memory, the always-on capability—without the existential dread of wondering what your AI is doing right now.

What OpenClaw Actually Does (And Why People Want It)

OpenClaw is an open source AI agent that runs continuously on your machine. Not "runs when you open it" or "runs when you ask it to"—it just runs. 24/7. You connect it to an AI model and interact through apps you already use: WhatsApp, Telegram, Slack. It has persistent memory that learns your patterns across sessions, and it handles automations and tasks while you sleep.

The appeal is obvious. Who wouldn't want a digital assistant that actually knows you and works autonomously? According to Snapper AI's breakdown of the situation, OpenClaw's popularity comes down to two core capabilities: it remembers everything about how you work, and it never stops working.

The problem is also obvious: it has access to your private data, gets exposed to untrusted content, and takes actions on your behalf with essentially no guardrails. As Snapper AI notes, "They are shipping security updates regularly, but at the moment, it's probably not quite there."

Claude's Controlled Alternative

Anthropic's response has been rolling out all week, and it's architecturally interesting. Instead of giving users a free-for-all autonomous agent, they're shipping the individual features that make OpenClaw compelling—but each one comes wrapped in safety protocols.

Scheduled Tasks landed in Claude's Cowork feature, which is the agentic side of Claude desktop. You can now set up recurring automations: a morning brief at 8 AM, weekly spreadsheet updates, Friday team summaries. The interface lets you configure task names, descriptions, prompts, model selection, working folders, and frequency (hourly, daily, or weekly).

The catch: these tasks only run while your computer is awake and Claude desktop is open. If your machine is asleep when a task is scheduled, Cowork skips it and runs it automatically once you're back online, sending you a notification about what it missed.

Snapper AI frames this as a deliberate trade-off: "Anthropic is deliberately trading that always on capability for safety. Your tasks run in a sandboxed environment. Claude asks for permission before doing anything destructive, and nothing happens without the app being open."

Auto-Memory addresses OpenClaw's other killer feature—persistent context across sessions. Previously, every time you started a new Claude Code session, you had to re-explain your project setup, preferences, and structure. Now Claude Code automatically saves useful context as it works: project patterns, key commands, debugging approaches, your preferred workflows.

The system uses two files: claude.md as your instructions file, and memory.md as Claude's memory scratchpad that it updates automatically. If you explicitly ask Claude to remember something, it writes it there. But it also learns and adds things on its own as it observes your behavior.

The feature just dropped and is still rolling out, but the X announcement already has 1.5 million views—suggesting there's real demand for AI agents that actually remember you.

Remote Control rounds out the updates by letting you start a session on your desktop machine and check on it from your phone, browser, or any other device. Everything still runs locally; the remote interfaces are just windows into your local session. It's currently in research preview, but it addresses the "work from anywhere" flexibility that OpenClaw provides.

The Architecture of Trust

What's emerging here is two fundamentally different philosophies about AI agents. OpenClaw gives you everything with no restrictions—you're responsible for what happens. Claude ships these features incrementally with guardrails at every layer: sandboxed environments, permission prompts before destructive actions, scheduled tasks that only run when your app is open.

Neither approach is inherently wrong. They're optimizing for different priorities.

If you need a fully autonomous agent running 24/7 and you're comfortable managing the security implications, OpenClaw is currently well ahead. The feature set is more complete, the autonomy is unrestricted, and the community is actively building around it.

But if you want automation without the risk of your inbox getting nuked or your files getting reorganized into oblivion while you're asleep, Claude's approach is increasingly viable. Snapper AI points out that Anthropic is essentially building "an AI agent that remembers you, automates recurring work, and runs without you babysitting it"—just with safety protocols that actually stop the agent when things go sideways.

The question isn't which system is "better." It's which set of trade-offs matches what you're actually trying to do. Because the demand for always-on AI agents is clearly there—that Meta researcher wanted her inbox cleaned up, she just didn't want it nuked. The difference between those two outcomes is architecture.

Anthropic is betting that most people will choose the version where the AI asks before it deletes everything. We're about to find out if they're right.

—Zara Chen, Tech & Politics Correspondent