Claude Code: Everything You Need to Know

From first prompt to agent teams — one guide.

A practical guide to Claude Code — from your first prompt to multi-agent automation, hooks, MCP, and team workflows. Built around clear mental models and real examples, not marketing.

npm install -g @anthropic-ai/claude-code

Who this is for: Developers using (or about to use) Claude Code. Beginners get a guided path; power users get depth on Skills, Hooks, MCP, and Agent Teams.

---

🧭 Choose your path

You are…	Start here	Time
🚀 New to Claude Code	Setup → Prompt Engineering → Your First Skill	~15 min
⚡ Already using it, want depth	Skills · Hooks · MCP	~30 min each
🧠 Building teams or automation	Agent Teams · Workflows & automation · BMAD	varies

---

🧠 When to use what

The four extension points in Claude Code, side by side:

Tool	Use when…	Skip if…	Lives in
Skills (slash commands)	You repeat the same prompt or workflow ≥3 times	One-off task	.claude/commands/*.md
Hooks	You want code to run automatically on tool use, session start, etc.	You only want manual triggers	.claude/settings.json
Subagents	A subtask is big enough to need its own isolated context	The task fits in your main session	.claude/agents/*.md
MCP servers	You need Claude to use external tools (browsers, DBs, APIs)	All your data is in local files	Configured per project

💡 These four compose. Most polished workflows combine 2–3.

---

📚 What's inside

Fundamentals — What is Claude Code? · Setup · Prompt Engineering

Workflow extensions — Slash Commands · Skills · Hooks

Multi-agent & integration — Subagents · Agent Teams · Workflows & automation · MCP

Productivity & frameworks — Effort levels · Fast Mode · Super Claude · BMAD Method

Reference — Slash Command Cheatsheet · Effort levels (full guide) · FAQ · Updates & Deprecations · Further Reading

What is Claude Code?

Claude Code is Anthropic's official CLI for working with Claude from your terminal. You point it at a project; it reads the code, plans, edits files, runs commands, and commits — all from the prompt line.

Three things it does that a chat UI can't:

• Reads your actual repo — not pasted snippets. Claude sees your file tree, runs grep, follows imports, and grounds answers in real context.
• Edits in place and runs your tests — diff-aware edits, then pytest/vitest/go test on the spot to verify the change.
• Composes with the rest of your stack — slash commands, hooks, sub-agents, MCP servers, and your normal git/shell workflow.

If you've used Copilot or Cursor, think of Claude Code as their "agent in your terminal" peer — same idea, different surface, no editor lock-in.

claude          # start a session in the current repo
> explain what this codebase does
> fix the failing test in src/api.test.ts
> open a PR with the changes

---

The Claude 5 era: today's model lineup

Three launches landed in quick succession this summer: Claude Opus 4.8 (May 28, 2026) took over as the Opus-tier flagship, Claude Fable 5 and its restricted sibling Claude Mythos 5 (June 9, 2026) opened a new Mythos-class tier above Opus, and Claude Sonnet 5 (June 30, 2026) became Claude Code's default model. 1M-token context is now standard across current Opus, Sonnet, and Fable models — no beta flag, no long-context surcharge — with 128K max output.

Choosing a model — quick guide:

Model	Reach for it when…
Sonnet 5 (default)	Everyday coding — most tasks live here. Intro pricing $2/$10 per MTok through Aug 31, 2026 (then $3/$15)
Opus 4.8	Complex reasoning, large refactors, orchestrating agents — $5/$25, unchanged from 4.7
Fable 5	Genuinely hard problems — Mythos-class capability above Opus at $10/$50
Haiku 4.5	Fast, lightweight tasks — quick questions, doc updates ($1/$5, 200K context)

Opus 4.7 / 4.6 and Sonnet 4.6 are now legacy models (still available via API and /model); Opus 4.1 retires August 5, 2026. Mythos 5 is the same underlying model as Fable 5 with fewer safeguards — invitation-only for approved organizations via Project Glasswing.

→ Full specs, capabilities, and pricing in docs/reference/models.md

---

Claude Code Setup

⏱️ 5-minute setup. Get from zero to your first AI-assisted commit.

1. Install

npm install -g @anthropic-ai/claude-code

Requires Node.js 18+. For other install methods (Homebrew, curl, native binary), see the official install guide.

2. Authenticate

claude

On first run, Claude Code opens a browser to sign in with your Anthropic account (Pro, Max, or API key all work). After that, you can re-authenticate any time with /login (and sign out with /logout) inside a session, or claude auth login|status|logout from your shell.

3. Run your first prompt

From any project directory:

cd ~/your-project
claude

Once Claude Code is running, try one of these:

• explain what this codebase does — Claude reads your repo and summarizes.
• add a README section about installation — generates content based on your project.
• find and fix the failing test in src/api.test.ts — diagnoses and edits in place.

4. (Optional) Generate a CLAUDE.md

/init

Creates a project-level instruction file that Claude reads on every session — your project's "house rules." More on this in Prompt Engineering Deep Dive.

5. (Bonus) See a real Claude Code project setup

This repo's own .claude/ directory is a working example of a fully-configured Claude Code project. Browse it as a reference for what a polished setup looks like:

Path	What it does
.claude/settings.json	Project-level Claude Code settings — permissions and hooks
.claude/agents/	5 specialized subagents (frontend, tech lead, PM, UX designer, code reviewer)
.claude/commands/	7 custom skills — /pr, /review, /tdd, /test, /five, /ux, /todo. Slash commands and Agent Skills are now one system — see Skills.
.claude/hooks/	Python hook scripts (post_tool_use.py, notification.py, stop.py, subagent_stop.py) — see Hooks

💡 Next: Once you're comfortable with the basics, jump to Claude Skills to build reusable slash commands in 3 minutes.

---

Prompt Engineering Deep Dive

📖 Claude Initialization Run the /init command to automatically generate a CLAUDE.md file. Your CLAUDE.md files become part of Claude's prompts, so they should be refined like any frequently used prompt. A common mistake is adding extensive content without iterating on its effectiveness. Take time to experiment and determine what produces the best instruction following from the model.

1. Explore → Plan → Code → Commit

Versatile workflow for complex problems.

• Explore: Read relevant files/images/URLs; use subagents for verification. Do not code yet.
• Plan: Ask Claude to make a plan. Use "think", "think hard", "think harder", or "ultrathink" to nudge depth in the prompt — see Effort levels for the full reasoning dial. Optionally save the plan for future reference.
• Code: Implement the solution; verify reasonableness as you go.
• Commit: Commit results, create pull requests, update READMEs/changelogs.
• Claude has two default modes: Plan Mode and Accept Edits Mode. You can toggle between them using the Shift + Tab keys.
  • 
  • 

💡 Pro Tip: Research & planning first significantly improves performance for complex tasks.

---

2. Test-Driven Workflow (Write Tests → Code → Commit)

Ideal for changes verifiable with unit/integration tests.

• Write Tests: Create tests based on expected inputs/outputs; mark as TDD.
• Run & Fail Tests: Confirm they fail; no implementation yet.
• Commit Tests: Commit once satisfied.
• Write Code: Implement code to pass tests; iterate with verification via subagents.
• Commit Code: Final commit after all tests pass.

🔹 Clear targets (tests, mocks) improve iteration efficiency.

---

3. Visual Iteration (Code → Screenshot → Iterate → Commit)

• Provide screenshots or visual mocks.
• Implement code, take screenshots, iterate until outputs match mock.
• Commit once satisfied.

🔹 Iteration significantly improves output quality (2-3 rounds usually enough).

---

4. Effort levels — how hard Claude thinks

→ Full guide in docs/reference/effort-levels.md

Mental model: Effort is a behavioural dial, not a token budget — it shifts thinking depth, tool-call appetite, response length, and how persistently Claude pushes through multi-step work. Higher ≠ smarter; context quality often matters more.

The API knows 5 levels (low → max, default high); Claude Code adds a sixth:

Level	Reach for it when…
low	Fast interactive queries you're steering — file renames, simple greps
medium	General coding, small refactors, autonomous sessions where the plan is clear
high	Multi-file refactors, complex debugging — the default on current models
xhigh	Long autonomous agentic sessions (Fable 5, Mythos 5, Opus 4.8/4.7, Sonnet 5)
max	Architecture, subtle bugs, security review — genuinely hard problems only. Session-only
ultracode (Claude Code only)	xhigh reasoning plus automatic multi-agent workflow orchestration. Session-only

Current defaults (July 2026): Opus 4.8 → high on all surfaces; Sonnet 5 → high on API and Claude Code. Check yours with /effort. (Historical footnote: Claude Code v2.1.117, April 2026, first standardized Pro/Max defaults to high after the March "nerfed medium" episode.)

Setting it, in order of persistence:

# This turn only — adds an in-context cue (does not change API effort)
> ultrathink — design the migration strategy

# This session — slider with no args, level name with arg
/effort xhigh
/effort ultracode               # xhigh + automatic multi-agent workflows
/effort auto                    # reset to model default

# All sessions (low/medium/high/xhigh) — add this key to .claude/settings.json:
#   "effortLevel": "high"
# max and ultracode are session-only by design and can't be persisted.

⚠️ Two gotchas worth knowing:

• max shows diminishing returns on routine work and is more prone to overthinking — Anthropic's own guidance. Don't default to it.
• Context quality often beats more effort. If you're reaching for max on a task that shouldn't need it, ~80% of the time the fix is upstream — sharper CLAUDE.md, atomic plan, named files. Full breakdown →

💡 Pattern: plan-with-Opus / execute-with-Sonnet. Plan in Opus 4.8 (or Fable 5) at xhigh or max; hand the atomic, zero-ambiguity plan to Sonnet 5 at lower effort to execute. Sonnet follows clear plans without drift, so the cheap execution is reliable when the plan is sharp.

---

Claude Commands

Claude Code ships dozens of built-in slash commands (official reference) plus the ability to define your own as skills (markdown files in .claude/commands/). The two work together — built-ins for common operations, custom skills for your team's workflows.

Day-1 essentials

Command	Purpose
/init	Generate a CLAUDE.md for your project — your "house rules" Claude reads every session
/help	List all available commands
/clear	Reset conversation history when you want a clean slate
/usage	Track token and plan usage (merged /cost + /stats as of v2.1.118)
/model	Switch models — your pick persists as the default for new sessions (press s for session-only)

→ Curated slash-command cheatsheet in docs/reference/commands.md (including /fast, /hooks, /mcp, /teleport, /workflows, /rewind, …)

Custom slash commands

Define a frequently-used prompt once as a markdown file, invoke it forever with /skill-name:

mkdir -p .claude/commands
echo "Analyze this code for performance issues and suggest optimizations:" \
  > .claude/commands/optimize.md

💡 Next level: custom slash commands and Skills are the same thing. Head to Claude Skills for the deep dive — built-in skills, the 7 custom skills in this repo, workflow recipes, and how to write your own.

---

Claude Skills

~3 min read · Full guide in docs/skills.md →

Mental model: Skills package a workflow into a markdown file. Two equivalent formats — officially one system now:

• Slash skills — .claude/commands/<name>.md, you invoke them with /<name>
• Agent Skills — .claude/skills/<name>/SKILL.md with YAML frontmatter; Claude can also auto-invoke these when the description matches the task

.claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy. Skills follow the open agentskills.io standard, adopted by ~40 products beyond Claude Code (Codex, Copilot, Cursor, Gemini CLI, …).

⚠️ Security: Skills are executable instructions running with your shell permissions. Read every third-party skill before adding it — exactly like reviewing a shell script before sourcing it.

Your first skill in 3 minutes

mkdir -p .claude/commands

cat > .claude/commands/analyze.md << 'EOF'
# Code Analysis

Analyze the current code for:
- Potential bugs and edge cases
- Performance optimizations
- Code quality improvements
- Security vulnerabilities

Provide specific, actionable recommendations.
EOF

claude       # then type: /analyze

That's it — a working slash skill. Promote it to an Agent Skill later by moving it to .claude/skills/analyze/SKILL.md and adding name/description frontmatter.

Want more depth?

The full Skills guide in docs/skills.md covers:

• The 7 custom slash skills in this repo's .claude/commands/: /pr, /review, /tdd, /test, /five, /ux, /todo
• Bundled built-in skills (e.g. /dataviz, /debug, /keybindings-help)
• Slash skills vs Agent Skills — when to use each, frontmatter contract, conversion path
• Workflow recipes — feature dev with TDD + PR, bug investigation, UX-first dev
• How to write your own skills (file format, scope, examples)
• Skills FAQ, troubleshooting, and best practices

Beyond your own skills — the ecosystem

The community has built an enormous catalog of Agent Skills. Three places to start browsing:

Resource	What it offers
anthropics/skills	Anthropic's official skills — PDF, slides, brand guidelines, document creation (158k+ ⭐)
SkillHub · SkillsMP · Smithery · skills.sh	Searchable marketplaces — community Agent Skills indexed from GitHub at massive scale
travisvn/awesome-claude-skills · ComposioHQ/awesome-claude-skills	Curated lists for high-signal picks

Notable community skills: skill-creator, skill-installer, mcp-builder, systematic-debugging, pair-programming, github-code-review, pptx, react, frontend-design, prompt-engineering-patterns, superpowers, brainstorming, market-research-reports, senior-data-engineer, and many more — see the full ecosystem section in docs/skills.md for categorized tables and install paths.

---

Hooks

Mental model: Hooks are programmable checkpoints on Claude Code's lifecycle (before/after a tool call, session start, prompt submit, etc.). Your script inspects the proposed action and returns allow / deny / modify.

Three cases that win most teams over:

Use case	What the hook does
Auto-format on save	Runs prettier / ruff / gofmt after every Edit so Claude's output matches your style
Block sensitive paths	Refuses changes to .env, secrets/, infra/prod/ regardless of what Claude tries
Action audit log	Records every tool call to a file — paper trail of what Claude did and when

If none of those resonate, skip ahead.

Setting up hooks

Hooks live in settings files at four scopes (later overrides earlier):

Scope	Path
User-wide	~/.claude/settings.json
Project (committed)	.claude/settings.json
Project (local, gitignored)	.claude/settings.local.json
Enterprise managed policy	platform-specific

Quickest setup — use the interactive menu:

/hooks    # browse, enable, configure hooks without touching JSON

Manual setup — for the hook scripts in this repo:

1. Copy .claude/hooks/ into your project's .claude/ folder.
2. Delete the hook scripts you don't need; keep the rest.
3. Install uv (required to run the Python hook scripts).
4. Copy .claude/settings.json into your project's .claude/ folder.
5. In settings.json, replace any hardcoded uv path with the output of $(which uv).

project-root/
└── .claude/
    ├── hooks/
    │   ├── notification.py
    │   ├── post_tool_use.py
    │   └── ...
    └── settings.json

Hook Events

Hooks run in response to various events within Claude Code's lifecycle: examples

• PreToolUse: Runs after Claude creates tool parameters but before processing the tool call.
• PostToolUse: Runs immediately after a tool completes successfully.
• Notification: Runs when Claude Code sends notifications, such as when permission is needed to use a tool or when prompt input has been idle.
• UserPromptSubmit: Runs when the user submits a prompt, before Claude processes it.
• Stop: Runs when the main Claude Code agent has finished responding (does not run if stopped by user interrupt).
• SubagentStop: Runs when a Claude Code subagent (Task tool call) has finished responding.
• SessionEnd: Runs when a Claude Code session ends.
• PreCompact: Runs before Claude Code is about to run a compact operation.
• SessionStart: Runs when Claude Code starts a new session or resumes an existing session.
• TeammateIdle: Runs when an agent teammate becomes idle (Agent Teams) — exit code 2 sends the teammate back to work.
• TaskCompleted: Runs when a task is marked as completed — exit code 2 blocks the completion.

These are the most-used events. The full catalog is 30 events (SubagentStart, PermissionRequest, FileChanged, WorktreeCreate, PostCompact, …) — see the official hooks reference.

Hook input

Hooks receive JSON via stdin. Every event includes session_id, transcript_path, and cwd. Event-specific fields:

Hook Event	Event-specific fields
PreToolUse	tool_name, tool_input
PostToolUse	tool_name, tool_input, tool_response
Notification	message
UserPromptSubmit	prompt
Stop / SubagentStop	stop_hook_active
PreCompact	trigger, custom_instructions
SessionStart	source
SessionEnd	reason
TeammateIdle	teammate_id, last_activity
TaskCompleted	task_id, task_name, completion_time

ℹ️ The team_name field in TaskCreated / TaskCompleted / TeammateIdle payloads is deprecated since v2.1.178 (one implicit team per session).

Hook output

Two ways to communicate back: exit codes for simple control, JSON in stdout for fine-grained behavior.

Exit code	Effect
0 (success)	stdout shown in transcript mode (CTRL-R). For UserPromptSubmit / SessionStart, stdout is added to Claude's context.
2 (blocking)	stderr fed back to Claude (or shown to user) to block the action. Stops tool calls in PreToolUse; stops prompt processing in UserPromptSubmit.
Other	stderr shown; execution continues.

Advanced: structured JSON in stdout. Per-event decision fields:

Event	JSON output
PreToolUse	permissionDecision: "allow" / "deny" / "ask"; updatedInput to modify tool parameters
PostToolUse	decision: "block" or undefined; additionalContext can be returned
UserPromptSubmit	decision: "block" or undefined; additionalContext can be returned
Stop / SubagentStop	decision: "block" or undefined
SessionStart	additionalContext

Security considerations

Hooks run arbitrary shell commands automatically with your user permissions — they can read, modify, or delete any file you can. Anthropic provides no warranty for what your hooks do.

Best practices:

• Validate and sanitize all inputs from stdin JSON
• Quote shell variables ("$var", not $var)
• Block path traversal (.., absolute paths outside the project)
• Use absolute paths for invoked scripts so PATH attacks don't redirect
• Explicitly skip sensitive files (.env, .git/, secrets/)

Claude Code snapshots your hook configuration at session start and warns if hooks change mid-session — review before applying.

Execution & debugging

• Timeout — defaults vary by hook type: 600s for command/http/mcp_tool hooks, 30s for prompt hooks, 60s for agent hooks (some events lower these — e.g. UserPromptSubmit command hooks get 30s). Configurable per hook.
• Parallelization — all matching hooks run in parallel; identical handlers are deduplicated automatically.
• Environment — hooks run in the current dir with Claude Code's env; CLAUDE_PROJECT_DIR is available.
• Debug — /hooks shows current config; claude --debug shows hook execution logs; test scripts manually with the JSON payload piped to stdin.

---

Subagents

Three building blocks for going beyond a single Claude session:

Building block	What it gives you	When to reach for it
Git worktrees	Multiple branches checked out simultaneously, each in its own folder + Claude session	You want to work on feature-A while Claude finishes feature-B
General-purpose subagents	Spawn isolated Claude sub-sessions from your main session for parallel sub-tasks	A task is large or context-heavy enough that the main session shouldn't carry it
Specialized subagents	Pre-written role prompts (security-reviewer, frontend-engineer, etc.) you drop into .claude/agents/	You want focused expertise without writing the role prompt yourself

1. Git worktrees — parallel branches, parallel sessions

Git worktrees let one repo have multiple branches checked out at the same time, each in its own folder. Pair them with one Claude Code session per worktree to run independent streams of work.

git worktree add -b feature-a ../feature-a    # create the worktree
cd ../feature-a && claude                     # start Claude in it
# Repeat in another terminal for feature-b. Each session is independent.
git worktree remove ../feature-a              # clean up when done

💡 Use tmux to keep each worktree's session attached even when you close the terminal.

2. General-purpose subagents — when one Claude isn't enough

From your main session, ask Claude to spawn subagents for a parallel sub-task. Each subagent runs in its own context window and reports a summary back, so the main session stays focused.

Analyze the implementation of the payment feature.
Spawn 5 subagents to accelerate the work.
Ultrathink.

3. Specialized subagents — drop-in role prompts

Specialized agents are pre-written role prompts you drop into .claude/agents/. Each one carries its own focus and tooling, so a security-reviewer agent stays focused on threats instead of also opining on code style.

An agent is just a markdown file with YAML frontmatter — create .claude/agents/security-reviewer.md (or ask Claude to write it; the old /agents wizard was removed in v2.1.198):

---
name: security-reviewer
description: Use after changes touching auth, input handling, or secrets — reviews diffs for vulnerabilities.
tools: Read, Grep, Glob
---

You are a security engineer. Review the changes for injection risks,
secrets in code, authZ/authN gaps, and unsafe input handling. Report
findings by severity with concrete fixes.

The description is what the main session uses to decide when to delegate — keep it to a couple of sentences of when-to-use criteria; the body below the frontmatter is the agent's system prompt. This repo's .claude/agents/ holds five working examples.

This repo ships 10 production-ready specialist prompts you can drop into .claude/agents/:

Role	System prompt	Role description
Backend Engineer	prompt	description
Frontend Engineer	prompt	description
Database Engineer	prompt	description
Tech Lead	prompt	description
Code Reviewer	prompt	description
Security Reviewer	prompt	description
UX Engineer	prompt	description
Design Reviewer	prompt	—
Project Manager	prompt	description
Business Analyst	prompt	description

📐 The agent-orchestration-workflow.canvas flowchart in specialized-agents/ visualizes how the role prompts compose into a pipeline. Open in Obsidian or any canvas-compatible viewer.

Orchestrating specialists from the main session

Once you have specialized agents, the main session orchestrates them:

Have backend-engineer suggest UI improvements; have frontend-engineer
implement them; have code-reviewer review the changes; have
frontend-engineer address the review feedback.

---

Agent Teams (Experimental)

Agent Teams is an experimental feature that lets a single Claude Code session coordinate multiple specialist agents through a shared task list. The main session acts as the team lead; teammates work on their tasks (sometimes in parallel), report progress, and update the shared list. Reach for it on full-stack features, large refactors, or anything where multiple perspectives genuinely help. Skip it for single-file edits and quick fixes.

Enable it

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1   # add to ~/.zshrc to persist
claude

How a team is structured

• Team lead (your main session) — assigns tasks, sets dependencies, reviews completed work.
• Teammates (specialist sub-agents) — focus on a single task each, update the shared list, can request help.
• Shared task list — single source of truth visible to everyone, tracks dependencies and status (pending / in progress / completed).

Mechanics as of mid-2026: each session has one implicit team — the old TeamCreate/TeamDelete tools were removed in v2.1.178, and teammates are spawned through the Agent tool's name parameter. Teammates display in-process by default; for split-pane views you need tmux or iTerm2 (teammateMode: "iterm2", v2.1.186+).

Multi-agent collaboration patterns

1. Parallel development — three independent streams, three teammates, one team lead synthesizing.

Task 1: frontend-engineer  – Build login UI
Task 2: backend-engineer   – Implement auth API
Task 3: qa-engineer        – Write integration tests

2. Sequential pipeline — one teammate's output feeds the next.

Task 1: backend-engineer   – Create database schema
  ↓ (blocks Task 2)
Task 2: backend-engineer   – Implement CRUD endpoints
  ↓ (blocks Task 3)
Task 3: frontend-engineer  – Build admin dashboard

3. Code review workflow — feature lands, multiple reviewers, then revision.

Task 1: feature-developer  – Implement new feature
  ↓ (completed)
Task 2: security-reviewer  – Check for vulnerabilities
Task 3: perf-reviewer      – Analyze optimization opportunities
  ↓ (both complete)
Task 4: feature-developer  – Address review feedback

4. Research → implementation — one analyst, multiple workers applying findings.

Task 1: research-agent     – Analyze existing codebase patterns
  ↓ (generates recommendations)
Task 2: implementation team – Apply findings across modules
  - Teammate A: auth module
  - Teammate B: payment module
  - Teammate C: notifications module

Example: full feature implementation

Team lead: coordinate overall strategy.

Teammates:
- planner: requirements → technical spec
- backend-dev: API endpoints
- frontend-dev: UI components
- db-specialist: schema design + migration
- qa-engineer: unit + integration tests
- security-reviewer: security audit

Dependencies:
1. planner → spec (blocks all)
2. db-specialist → schema (blocks backend-dev)
3. backend-dev + frontend-dev in parallel
4. qa-engineer waits for implementation
5. security-reviewer waits for all code

Best practices

✅ Do	❌ Don't
Give teammates clear, focused, single-purpose tasks	Spawn more than 3–5 teammates per session
Use descriptive names (frontend-specialist, not agent1)	Assign vague or overlapping tasks
Set explicit task dependencies (blockedBy)	Skip dependencies — they prevent merge conflicts
Let teammates work in parallel when possible	Micromanage; trust the role prompt
Watch progress in the claude agents dashboard	Mix different project contexts in one team

Common issues

Issue	Fix
Teammates conflicting on the same files	Add explicit blockedBy dependencies
Team too slow	Reduce team size; increase parallelization
Tasks stuck	Check for circular dependencies
Context drift	Make task descriptions and teammate names more specific

Limitations (it's experimental)

• Multiple agents consume more tokens — budget accordingly.
• Large teams add coordination overhead; 3–5 teammates is the sweet spot.
• One team per session, no nested teams, and the lead is fixed; in-process teammates can't be resumed after a restart.
• Deeper models make better leads — Opus 4.8 or Fable 5 as lead, Sonnet 5 teammates is a solid split.

📚 Authoritative reference: code.claude.com/docs/en/agent-teams.

---

Beyond one terminal — the 2026 automation surface

Claude Code grew a set of orchestration features in mid-2026 that compose with everything above:

Feature	What it does	Docs
Dynamic workflows	Claude writes and runs an orchestration script that fans out tens to hundreds of subagents in the background; watch runs with /workflows. Opt in per session with /effort ultracode	workflows
Cloud code review	/code-review ultra runs a multi-agent review in the cloud (alias /ultrareview — 3 free runs on Pro/Max, then usage credits); claude ultrareview runs it non-interactively for CI	commands
Routines	/schedule (alias /routines) runs scheduled agents on Anthropic-managed cloud infrastructure; /loop and the Cron tools cover local scheduling	scheduled tasks
Artifacts (beta)	Publish live, shareable web pages to claude.ai