/loop
Anthropic’s native Claude Code command that gives agents a heartbeat — the ability to run on a schedule without human interaction. Shipped in March 2026. One of the three essential primitives for building agents (alongside memory and tools).
What It Does
/loop lets Claude Code run a job on a repeating cadence without you being at the computer. The agent wakes up, executes its task, and goes back to sleep. Before /loop, achieving this required external scaffolding: bash scripts, markdown context files, restart logic, and eval conditions “bolted on with duct tape.”
Why It Matters
/loop is the proactivity primitive that completes the OpenClaw-equivalent stack natively within Claude Code:
| Before /loop | After /loop |
|---|---|
| Agent only moves when you push it | Agent acts on its own schedule |
| You are the metronome | Agent has its own heartbeat |
| External scaffolding required | Native Anthropic command |
Combined with a memory system (like OpenBrain) and MCP tools, /loop enables agents that accumulate value across cycles:
- Loop alone = parrot that repeats advice each time
- Loop + memory = detective that builds a case across weeks of data
“The value of a loop isn’t in any single cycle. It’s in the accumulation across cycles.” — Nate B Jones
Common Patterns
- Morning check-in: Agent wakes at 6am, reads memory, pattern-matches against history, gives informed recommendations
- Overnight coding loop: Agent works on test coverage / refactoring / migration while you sleep, remembering what it tried
- Sales pipeline: Agent reviews leads, cross-references history, drafts outreach using best-performing messaging from memory
- Content calendar: Agent checks scheduled posts, searches web for relevant breaking news, flags stale references
Current Limitations (as of March 2026)
- No built-in “done” signal: Runs until expiration or manual stop. For goal-directed work, agent must check memory for completion state or run evals.
- Session-scoped: Closes when laptop closes. Not an always-on daemon.
- Terminal-first: Only available in Claude Code (CLI). Eventually expected to reach more user-friendly interfaces like co-work and claude.ai.
See Also
- OpenBrain — the memory primitive; combined with /loop, enables pattern-matching across cycles
- OpenClaw — the framework /loop replicates without the security risks
- Claude Code — where /loop lives
- AutoResearch and Evals — methodology for goal-directed loops with objective metrics
- Nate B Jones — source
- loop + OpenBrain