Background agents in Claude Code: the map of four ways to stop watching the terminal

TL;DR There are four mechanisms in Claude Code for letting Claude keep working without you watching: Agent View (claude agents or claude --bg — local persistent sessions managed by a supervisor independent of your terminal), headless mode (claude -p — single-shot, ideal for CI/CD and scripts), /loop (in-session, repeats a prompt at intervals inside an active session), and Routines (cloud at claude.ai, recurring or triggered by schedule/API/GitHub). Decision rule: does it run on your machine or in the cloud? Once or recurring? Started interactively or fired by an external event? Confusing them is the main source of wasted time in "autonomous" setups.

"Background agent" isn't one feature in Claude Code — it's a category with four tools that overlap in purpose and diverge in mechanism. People Google "claude code background agent" and land on Anthropic's autonomy blog post, but that page doesn't compare: it just announces. The practical question — which of the four do I need today — goes unanswered.

What they share: all four let Claude keep working without continuous input from you. What separates them: where they run, how long they live, what triggers them, and which quota they consume.

Result:

> claude agents

Pinned
  ✽ payment-retry-fix      Edit src/payments/retry.ts            3m

Ready for review
  ∙ stripe-webhook         github.com/acme/api/pull/2048      ●  2h

Needs input
  ✻ feature-flag-cleanup   needs input: rollout 50% or 100%?     1m

Working
  ✽ flaky-test-investig    run 12 · all checkpoints cleared      4m
  ✢ nightly-deps-check     /loop · next run in 18m

Each row is a full Claude Code session running in the background. Close the terminal, come back tomorrow, they're still there.

The four mechanisms

1. Agent View (claude agents) — persistent local sessions

claude agents
# or create one without opening the view:
claude --bg "investigate the flaky SettingsChangeDetector test"
# or background the current session:
/bg

A local supervisor keeps N sessions alive independently of your terminal. Each session is a full Claude Code conversation with its skills, hooks, MCPs, permissions. Close your shell — they survive. Reboot the machine — they fail, but you can attach to them again.

It's the canonical entry point to "background". Covered in depth in Agent View: one screen for every agent you've dispatched. If what you need is to see what's alive inside one session (subagents, monitors, background commands), that's the Tasks panel — they complement: Agent View shows sessions; Tasks shows the tasks inside each one.

When: parallel local work (PR review + bug fix + investigation + refactor). Edits isolated in .claude/worktrees/ automatically — pair this with worktrees for parallel tasks and you get per-session isolation without stepping on your main branch.

2. Headless mode (claude -p) — single-shot for CI and scripts

claude -p "explain the root cause of this error" --bare < build.log

Non-interactive, one agent turn. Reads stdin, writes stdout, exits with a status code. It's the basis of the Agent SDK and of Claude Code in GitHub Actions. Covered in headless mode as autonomous agent.

⚠️ Critical pricing caveat: starting June 15, 2026, claude -p on subscription plans draws from a separate monthly Agent SDK credit distinct from interactive usage. If your CI runs hot, plan accordingly.

When: CI/CD, build scripts, package.json hooks, Claude-as-linter, pipes from external tooling.

3. /loop — in-session recurrence

> /loop 1h /goal Verify the design system still matches the latest
  Figma export pixel-perfect and ping me if any diverged.

Repeats a prompt every interval (m/h/d) inside an active session. Shows up as a "Working/sleeping" row in Agent View. Useful for continuous monitoring or periodic checks during a long session. Covered in /loop for recurring tasks. The most common combo — /loop with /goal as a stop condition on each iteration — lives exemplified in the pixel-perfect Figma + Chrome MCP pipeline. If you want to watch events in real time while it runs, the Monitor tool is the natural complement.

When: checks during a long session (smoke tests during a refactor, dev server monitoring, polling a remote CI). Local — dies if you close the session.

4. Routines — cloud with triggers

> /schedule daily PR review at 9am

Routines run on claude.ai/code (cloud, not your machine). They fire on:

• Schedule (cron, hourly/daily/weekly, or one-off natural language)
• API (HTTP POST with bearer token, ideal for alerts and deploys)
• GitHub events (PR opened, release, etc.)

Your laptop doesn't need to be on. Covered in Routines: cloud agents with triggers. The full map of /schedule vs /loop vs cron lives in Schedule vs Loop vs Cron. If you need the routine to ping you via chat without opening the web, wire it through Channels and Telegram or remote control from your phone — those are the canonical paths for notification + intervention without a terminal.

When: recurring automation that must run whether or not you're there (nightly PR review, alert triage, docs drift), or external-system triggers (Sentry → routine, deploy → routine).

The decision table

The most common mistake

"I put /loop to review PRs every hour but it stopped when I closed my laptop."

/loop is local in-session. Close the session → it dies. What you wanted was a Routine (cloud + schedule trigger). The reverse pattern: putting in a Routine something that runs once a day for 10 seconds — overkill. Use claude --bg with a clear name.

The golden rule: does this need to survive turning off the machine? If yes, Routine. If no, the other three depending on the rest of the flowchart.

Quick reference

Official docs: Agent View · Run Claude Code programmatically (headless) · Routines · /loop and in-session scheduling

Requirements

• Agent View: Claude Code v2.1.139+
• --permission-mode/--effort on claude agents: v2.1.142+
• Routines: Pro/Max/Team/Enterprise plan with Claude Code on the web enabled
• Headless with separate Agent SDK credit: starting June 15, 2026