Claude code background agents

Claude Code Background Agents and Tasks

When to use: Background Bash via Ctrl+B or run_in_background=true, /loop with fixed or self-paced intervals, Monitor tool for streaming output, Agent view (claude agents) for parallel sessions, subagents for side tasks, workflows for multi-agent orchestration, /schedule (cloud routines), and /goal for turn-by-turn looping.

---

Quick Summary

Background Bash: Run a shell command asynchronously. Press Ctrl+B during a Bash tool call, or Claude sets run_in_background=true. Command runs; output written to file in session directory. Claude reads output with Read tool. Returns task ID. List and stop with /tasks.

/loop: Recurring prompt on a session-scoped schedule. Fixed interval (5m, 30s) or self-paced. Behind the scenes uses CronCreate/CronList/CronDelete tools. Expires 7 days or when you Esc.

Monitor: Built-in tool; streams background script output line-by-line as each line arrives. Claude reacts mid-conversation without polling. Not available on Bedrock/Vertex/Foundry. More token-efficient than /loop on fixed cron.

Agent view (claude agents): Dispatch and monitor independent background sessions in a dedicated terminal UI. Each session isolated, gets its own worktree. Survives terminal exit (supervisor process keeps them alive). You hand off work; Claude runs autonomously.

Subagents: Spawn a worker inside conversation with separate context. Foreground subagents show permission prompts as they run; background subagents don't and auto-deny unpermitted calls. Returns summary. Does not block main turn.

Workflows: JavaScript script runs in background, orchestrates many subagents. Script holds plan; intermediate results stay out of context. Resumable. Invoke via /workflows or ultracode effort level.

/schedule (cloud routines): Create a task on Anthropic infra. Cron or API trigger. Persists independently of any session.

/goal: Keep session looping turn-by-turn until a condition is met. Session stays active.

---

Background Bash Commands

When: Long builds, test suites, dev servers, file watches.

How to invoke:

1. Press Ctrl+B during a Bash tool call to move it to background.
2. Or: Claude sets run_in_background=true on the Bash tool parameter.

Behavior:

• Command runs asynchronously. Claude gets back task ID immediately.
• Output written to file in session directory.
• Claude reads output later with Read tool (no blocking).
• Max 30,000 chars per command before file save + preview. 5GB hard output cap before auto-termination.
• Auto-cleaned on session exit.
• List and stop: /tasks command.
• Disable: CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1.

Key constraint: Background Bash does NOT re-invoke Claude on completion. You or a scheduled /loop must poll or read the output file.

---

/loop: Local Recurring Prompts

Scope: Session-local. Expires 7 days or Esc. Restored on --resume if unexpired.

Three modes:

Mode 1: Fixed Interval + Prompt

/loop 5m check if deployment finished

• Interval: s, m, h, d. Converts to cron; Claude confirms actual schedule.
• Fires between turns, never mid-response. If Claude busy, waits for turn end.
• Jitter: up to 30 minutes after scheduled time (deterministic per task ID to prevent API thundering).

Mode 2: Prompt Only (Self-Paced)

/loop check whether CI passed and address review comments

• Claude chooses delay 1m–1h after each iteration based on observed work.
• May use Monitor tool instead to avoid re-polling.
• Respects 7-day expiry.
• Bedrock/Vertex/Foundry: unsupported; falls back to fixed 10m instead.

Mode 3: Bare /loop (Fixed Built-in Maintenance)

/loop
/loop 15m

• Bare /loop = built-in maintenance prompt on fixed cron (NOT self-paced). Customizable via .claude/loop.md or ~/.claude/loop.md.
• Bedrock/Vertex/Foundry: bare /loop prints usage instead.

Stop: Press Esc (works only during wait). Fixed-cron tasks keep running. Self-paced /loop may stop itself if work complete.

Under the hood: CronCreate, CronList, CronDelete tools. Max 50 tasks per session.

---

Monitor Tool

What: Streams background script output line-by-line without polling.

When: More token-efficient alternative to /loop fixed intervals. Claude may choose Monitor for self-paced /loop instead of cron.

How: Claude runs script in background; yields each output line as it arrives. Claude reacts to each line immediately.

Availability: Not on Bedrock/Vertex/Foundry.

---

Agent View: Background Sessions

Command: claude agents (opens dedicated terminal UI).

What: One screen for all dispatched background sessions. Each is a full Claude Code session, independent, with its own worktree.

How to dispatch:

• From agent view: type prompt, press Enter.
• From inside session: /background [prompt] or /bg [prompt].
• From shell: claude --bg "<prompt>".

Behavior:

• Each session isolated (edits go to separate worktree under .claude/worktrees/).
• You hand off; Claude runs autonomously.
• Status at a glance. Attach/peek/reply without leaving agent view.
• Supervisor process keeps sessions alive after terminal exit. Restart machine requires claude respawn <id> recovery.

---

Subagents (In-Session Delegation)

Scope: Inside one conversation. Spawned by Claude or via /agents command.

Permission behavior (critical distinction):

• Foreground subagents: Show permission prompts as they run (same as main session).
• Background subagents: Don't show prompts; inherit permissions already granted; auto-deny unpermitted calls and keep going.

Behavior:

• Runs in own context.
• Returns summary to main conversation.
• Does not accept new user input mid-task.
• Can inherit full tool access or restricted set.

When: Side research, exploration, parallel investigation.

---

Dynamic Workflows

What: JavaScript script orchestrates many subagents in background. Script holds plan; intermediate results stay out of context.

When: Large audits, 500-file migrations, cross-checked research, multi-angle planning.

How: Mention ultracode keyword or ask Claude to write workflow. Script executes in background. Monitor with /workflows. Final report to session when done.

Resumable: Pause/resume within same session. Exit Claude Code = fresh restart next session.

---

/schedule: Cloud Routines

Scope: Persists on Anthropic infra, independent of any session. Survives machine restart.

Triggers: Cron schedule, API call, GitHub event.

vs /loop: /loop session-local, fires only while Claude running locally. /schedule runs in cloud.

---

/goal: Turn-by-Turn Looping

What: Keep session actively looping until condition is met.

How: /goal all tests pass and code is deployed

Behavior: Claude evals condition after each turn. Session stays awake.

---

Comparison Table: When to Use

Use case	Tool	Scope	Blocking?	Survives exit?	Multi-agent?
Long build/test, check later	Background Bash	Session	No	No	No
Poll CI every 5m while working	/loop 5m	Session	No	7d	No
Stream real-time output	Monitor	Session	No	No	No
Dispatch independent task	Agent view	Session	No	Yes (supervisor)	Many
Side research, return summary	Subagent	Conversation	No	No	Few
Codebase audit, 100s of agents	Workflow	Session	No	7d	Many
Cron on cloud, no machine needed	/schedule	Cloud	No	Yes	No
Active loop until done	/goal	Session	Partially	No	No

---

Polling Without Re-Running Claude

Pattern: Bash loop + file writes → Claude reads file periodically.

while true; do
  curl https://api.example.com/status > /tmp/status.json
  sleep 30
done

Claude runs in background. Claude reads /tmp/status.json later. No re-invocation.

Better: Use Monitor or /loop if Claude should react, or /goal for tight feedback.

---

Key Constraints

1. Background Bash does not re-invoke Claude. Poll with /loop, Monitor, or read file.
2. Subagents: foreground shows prompts; background doesn't. Critical for permission handling.
3. Agent view sessions are discrete. Separate worktrees; don't see main conversation.
4. Scheduled tasks (cron) expire 7 days. Extend by recreating, or use /schedule for durable cloud tasks.
5. /loop fires between turns, never mid-response. If Claude busy, waits for turn end.
6. Jitter deterministic per task ID. Prevents API thundering at wall-clock moments.
7. Monitor not on Bedrock/Vertex/Foundry. Use /loop instead.

---

Common Patterns

Watch a deploy: /loop 2m did deployment finish? if so, run smoke tests

Babysit a PR: /loop check CI status and new review comments; address them

Audit entire codebase: ultracode: audit every function in src/ for missing error handling

Parallel research: Dispatch three independent investigations via agent view, review findings.

Continuous improvement: /goal all tests pass and code is deployed

---

Monitoring Active Work

• Background Bash: /tasks command shows running tasks.
• /loop: /tasks shows scheduled tasks.
• Agent view: claude agents opens UI.
• Subagents: /agents shows Running tab.
• Workflows: /workflows shows phases, agent counts, token totals.

---

Token Costs

• Background Bash: runs locally, no ongoing tokens.
• /loop fixed: one prompt per fire, cache reuse.
• /loop self-paced: one prompt per iteration, Claude chooses delay.
• Monitor: streaming; efficient vs polling.
• Agent view: each session fresh context, multiplies tokens.
• Subagents: separate context per agent.
• Workflows: many agents; estimate before large run.
• /schedule (cloud): billed separately, runs independently.