Vibe Coding for Production
Scaling automation

Parallel and scheduled work

Scale Claude Code with background sessions, worktrees, batch agents, goals, loops, and cloud routines.

Claude Code now has several ways to keep work moving. They differ in isolation, lifetime, and where execution happens; choose deliberately rather than calling all of them “background agents.”

Background and forked sessions

  • Press Ctrl+B to background the current task.
  • Use /tasks to inspect task state.
  • /background <prompt> (alias /bg) detaches a session.
  • /fork <prompt> creates an independent background copy.
  • /branch <prompt> branches the conversation and switches to that direction.
  • claude agents opens Agent view for background sessions.

These flows are useful for independent test runs, investigations, and long commands while you continue another thread. See Agent view.

Worktree isolation

Start a session in an isolated Git worktree:

claude --worktree feature-auth

The short alias is -w. Custom subagents can request the same boundary with isolation: worktree in frontmatter. Use .worktreeinclude for required gitignored inputs that should be copied into new worktrees.

Worktrees isolate files, not external systems. Two agents can still collide in the same database, cloud account, port, queue, or test environment. See Worktrees.

Batch work

For a large task that splits cleanly:

/batch migrate the independent API handlers to the new validation pattern and
run each handler's tests

/batch creates a plan and dispatches roughly 5–30 worktree-isolated subagents. Use it only when file ownership and verification can be partitioned; a tightly coupled refactor usually benefits from one coordinated session.

Agent teams (experimental)

Agent teams let teammates coordinate directly, but they are experimental, token-heavy, and not automatically isolated in separate worktrees.

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Use them for research or genuinely separable ownership, not as a default for every multi-file task. See Agent teams.

Continue until a goal

Set an explicit completion condition:

/goal Keep working until the unit tests and production build both pass.

Clear it with:

/goal clear

A goal continues turns toward a condition; it does not prove the condition. Require command output and independent checks. See Goal.

Loop, session schedule, or routine?

/loop 5m check the deployment and report only state changes

/loop repeats while the session stays active. Session scheduling uses CronCreate, CronList, and CronDelete and also depends on that session.

Routines are persistent cloud automation. Create or manage them with:

/schedule check the release pull request every weekday and summarize blockers
/schedule list
/schedule run

/routines is an alias. Routines can respond to schedules, API calls, and GitHub events without your terminal staying open. See Scheduled tasks and Routines.

Decision guide

NeedUse
Keep one local task runningCtrl+B or /background
Create a separate background copy/fork
Branch and switch conversational direction/branch
Isolate concurrent code changes--worktree
Split many independent code units/batch
Continue until a testable condition/goal
Poll while this session is open/loop
Persistent remote trigger or scheduleRoutine

On this page