Anthropic Just Gave Your AI Agent the One Thing OpenClaw Has. Without the Risk.
TL;DR
Anthropic’s little-noticed /loop feature is the missing “heartbeat” for agents — Nate argues that Claude Code now has the last core primitive OpenClaw popularized: the ability to wake up on a schedule and do work without you prompting it each time.
The agent recipe is simple: memory + proactivity + tools — His framework is that a real agent stops being a chatbot when it can read/write persistent memory, run on its own rhythm, and use tools to touch outside systems like APIs, databases, Slack, calendars, or Remotion.
You can get a lot of OpenClaw-like behavior without OpenClaw’s security mess — Instead of installing a risky framework, Nate says a SQL database behind an MCP server (“OpenBrain”), plus Claude Code /loop, plus tool access reproduces many of the same capabilities with less exposure.
Memory is what turns a proactive bot from a parrot into a detective — In his examples, an energy coach agent goes from generic daily advice to spotting that low energy happened 9 of the last 12 mornings and correlates it with late eating rather than caffeine.
The real payoff comes from compounding loops across time, not one-off runs — He connects /loop to overnight coding agents, sales pipeline triage, customer success monitoring, job search prep, and Andrej Karpathy’s Auto Research, where 100 remembered cycles outperform a single smart pass.
The catch: this is powerful but still early, technical, and session-bound — Nate is explicit that /loop has no built-in done signal, lives inside Claude Code, stops when the session ends, and still requires technical comfort, but he calls terminal use a kind of “time travel” because developers get these capabilities months ahead of mainstream UI users.
The Breakdown
From OpenBrain to a Full Agent
Nate starts by saying OpenBrain — his simple SQL database plus MCP server setup — gave AI memory, and thousands of people in his community actually built it. But memory alone isn’t enough: he now wants to give that memory a “heartbeat, a proactive rhythm,” plus tools, so it can do useful work instead of just storing facts.
Why /loop Matters More Than People Realized
The big news is Anthropic’s new /loop command in Claude Code, which most people ignored. Nate’s claim is blunt: /loop is the last missing piece needed to recreate much of what people wanted from OpenClaw, because it lets an agent wake up, check things, act, and go back to sleep without you playing human metronome.
The Three Lego Bricks: Memory, Proactivity, Tools
He reduces the whole agent story to three primitives. Memory keeps the agent from being “a new hire on their very first day,” proactivity means it can move without waiting for a prompt, and tools keep it from being “a brain in a jar” by letting it call APIs, write data, trigger workflows, and generate outputs.
The Energy-Tracking Example That Shows Why Memory Changes Everything
Nate walks through a simple personal example: an agent checks in every morning about sleep, food, and energy. Without memory, it just hands out generic advice; with a SQL-backed memory store, it can notice you were tired 9 of the last 12 mornings, compare the good days, and conclude late meals matter more than caffeine — a shift he sums up as the difference between “a parrot that repeats a phrase” and “a detective that builds a case.”
Business Use Cases: Customer Success, Networking, and Job Applications
He quickly shows the same pattern in work settings. A customer success agent can see a client’s usage fell 15% last week and 28% now, match that trajectory to a failed account from 6 months ago, and recommend executive outreach; a networking agent can pull context on people you’re about to meet and generate a personalized briefing video via Remotion; a job-search agent can update your cover letter with Tuesday’s deployment metric and Wednesday’s client testimonial before your Thursday application block.
The Bigger Point: Loops Compound Value Over Time
This is where Nate zooms out: the value of a loop isn’t one cycle, it’s what accumulates across cycles. He ties this to the overnight coding-agent trend, where developers already had agents iterating until tests passed, and says /loop plus memory turns what used to require external scaffolding, markdown handoffs, and duct-taped evals into something closer to actual architecture.
Karpathy’s Auto Research and the Case for Remembered Iteration
To make the point vivid, he cites Andrej Karpathy’s open-source Auto Research project, where an agent runs roughly 100 overnight experiments on a small model, keeping a persistent log of what worked and what didn’t. He says Shopify CEO Tobi Lütke used that repo to produce a smaller model that beat a larger hand-tuned one — not because any single run was genius, but because the system learned over many remembered cycles.
OpenClaw Without the Chaos
Nate ends by directly addressing the comparison: yes, he is basically describing OpenClaw-like agents, but “without the chaos.” He calls OpenClaw a security nightmare, notes Peter Steinberger himself has warned less technical users off it, and argues that native Claude scheduling, a database you control, and carefully scoped tools can deliver a lot of the same upside — while also admitting the current setup is still technical, session-limited, and missing built-in completion logic.