A quick reference to the ways you direct one or many agents in Claude Code, for the multi-agent build in this repo. Verified against Claude Code v2.1.x (June 2026).
The live source of truth is
/help. Type/help(or just/) in your session to see the exact commands your version supports — the set ships fast and varies by version. Treat this file as orientation, not gospel.
You don't need a command to send out agents. The primary way is plain English:
Build the server, web and mcp folders in parallel using separate subagents,
one per folder. Agree the API contract first and don't let them change it.
The commands below are for control and scale — naming agents, scoping them, backgrounding them, watching them, and setting an outcome to run toward.
| Command | What it does |
|---|---|
/goal [condition] |
Sets a completion condition; Claude keeps working across turns until it's met. With no argument, shows the current/last goal. |
/goal clear |
Removes an active goal early. Also accepts stop · off · reset · none · cancel. |
Pair /goal with Auto mode (Shift+Tab to cycle) and it runs unattended for hours.
A goal is only as good as its finish line — make it provable (the agent shows evidence in chat),
fenced (name the files, say what must not change), and guarded (branch + limits + review).
Goal: every test in /server passes — npm test exits 0. Don't edit any test file.
| Command / key | What it does |
|---|---|
/agents |
Manager UI: create, edit, scope and run named subagents. The Running tab lists live ones. |
@agent-<name> |
Force a specific named subagent onto the next task (guaranteed — not Claude's choice). |
/fork <directive> |
Spawns a forked subagent that inherits your full conversation, works in the background, and returns its result when done. |
Ctrl+B |
Send a currently-running subagent to the background and keep working. |
A subagent is just a markdown file with YAML frontmatter (.claude/agents/<name>.md):
---
name: backend
description: Owns /server — the SQLite schema, the DB, and the REST API
tools: Read, Edit, Write, Bash, Glob, Grep
model: sonnet
---
You own the /server folder only. Build to the agreed API contract.
Never touch /web or /mcp. Return the final endpoint list when done.Agents made via /agents work immediately; hand-edited files on disk need a session restart.
| Command / key | What it does |
|---|---|
/background [prompt] (alias /bg) |
Detaches the whole session to run as a background agent and frees your terminal. |
/tasks (alias /bashes) |
View and manage everything running in the background of the session. |
/batch <instruction> |
Decomposes a big codebase-wide change into 5–30 independent units, each in its own git worktree, run in parallel. |
\ (backslash) |
Opens Agent View — one dashboard listing every running/background session. |
claude agents (shell) |
Monitor/manage background agents from the terminal. |
claude --bg / --background (shell) |
Start a background agent straight from the shell. |
No slash command. Enable it, then ask in natural language. Use only when workers need to discuss or challenge each other (research, review, debugging competing hypotheses) — they cost far more tokens than subagents.
- Enable: set
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1insettings.jsonor your environment. - Start: "Spawn 3 teammates to review PR #142 — one on security, one on performance, one on tests."
- Flag:
claude --teammate-mode auto|tmux|iterm2|in-process(split panes need tmux or iTerm2). - Agent panel keys:
↑/↓select ·Enteropen & message ·xstop ·Escinterrupt ·Ctrl+Ttoggle shared task list.
| Command | What it does |
|---|---|
/autofix-pr [prompt] |
Spawns a Claude Code on the web session that watches your branch's PR and pushes fixes when CI fails or reviewers comment. |
/plan [description] (plan mode before agents act) · Shift+Tab (cycle plan/auto modes) ·
/model, /effort (tune reasoning) · /rewind (roll code + conversation back if an agent goes wrong) ·
/context + /compact (manage the window when it fills).
| Situation | Reach for |
|---|---|
| Everyday parallel work (this repo's build) | /agents + plain English |
| Peel off a side-task without losing context | /fork |
| Codebase-wide sweep (rename, migration) | /batch |
| "Run until it's green / done" | /goal + Auto mode |
| Workers need to argue with each other | Agent Teams |
| Fire-and-forget, come back to a PR | /background, cloud agents |
Split the PRD across subagents — /web, /server, /mcp (and an optional /test) — one folder
per owner, with the API contract agreed up front so they don't collide. That's section 1 of
BUILD-PROMPT.md. Use /plan to review before they write code; \ (Agent View) to watch them.
Sources: Claude Code docs — Commands,
Subagents,
Agent teams,
Agent view,
Goal. Current as of June 2026; run /help for your version.