Parallel Development with agent-cli dev

This skill teaches you how to spawn parallel AI coding agents in isolated git worktrees using the agent-cli dev command.

agent-cli dev supports two complementary patterns:

• Separate worktrees for isolated implementation/review tasks
• Multiple agents on the same worktree using dev agent --tmux-session <unique-name> for autonomous launches, or -m tmux when a human explicitly wants shared tmux windows

Installation

If agent-cli is not available, install it first:

# Install globally
uv tool install agent-cli

# Or run directly without installing
uvx agent-cli dev new <branch-name> --prompt "..."

When to spawn parallel agents

Spawn separate agents when:

• Multiple independent features/tasks can be worked on in parallel
• Tasks benefit from isolation (separate branches, no conflicts)
• Large refactoring that can be split by module/component
• Test-driven development (one agent for tests, one for implementation)

Do NOT spawn when:

• Tasks are small and sequential
• Tasks have tight dependencies requiring constant coordination
• The overhead of context switching exceeds the benefit

Core command

For new features (starts from origin/main):

agent-cli dev new <branch-name> --prompt "Implement the new feature..."

For work on current branch (review, test, fix) - use --from HEAD:

agent-cli dev new <branch-name> --from HEAD --prompt "Review/test/fix..."

For longer prompts (recommended for multi-line or complex instructions):

agent-cli dev new <branch-name> --from HEAD --prompt-file path/to/prompt.md

This creates:

1. A new git worktree with its own branch
2. Runs project setup (installs dependencies)
3. Saves your prompt to a unique task file in .claude/ in the worktree (for reference)
4. Opens a new terminal tab with an AI coding agent
5. Passes your prompt to the agent

Important: Use --prompt-file for prompts longer than a single line. The --prompt option passes text through the shell, which can cause issues with special characters (exclamation marks, dollar signs, backticks, quotes) in ZSH and other shells. Using --prompt-file avoids all shell quoting issues.

Automation rule

When an assistant is executing this workflow on the user's behalf, the spawn is not complete unless the agent receives a prompt at launch time.

• Prefer --prompt-file; create the prompt file first, then launch the agent
• Use dev new ... --prompt-file ... for a new delegated task
• Use dev agent ... --prompt-file ... for another agent in an existing worktree
• Do not stop after dev new ... alone if the user's intent was to delegate work immediately
• Do not run dev new ... --start-agent, dev new ... --agent <name>, or dev agent ... -m tmux without --prompt or --prompt-file unless the user explicitly wants an interactive session that they will drive manually

Writing effective prompts for spawned agents

Spawned agents work in isolation, so prompts must be self-contained. Include:

1. Clear task description: What to implement/fix/refactor
2. Relevant context: File locations, patterns to follow, constraints
3. Report request: Ask the agent to write conclusions to .claude/REPORT.md

Using --prompt-file (recommended)

For any prompt longer than a single sentence:

1. Write the prompt to a temporary file (e.g., .claude/spawn-prompt.md)
2. Use --prompt-file to pass it to the agent
3. The file can be deleted after spawning

Example workflow:

# 1. Write prompt to file
# 2. Spawn agent with the file
agent-cli dev new my-feature --prompt-file .claude/spawn-prompt.md
# 3. Optionally clean up
rm .claude/spawn-prompt.md

Prompt template

<Task description>

Context:
- <Key file locations>
- <Patterns to follow>
- <Constraints or requirements>

When complete, write a summary to .claude/REPORT.md including:
- What you implemented/changed
- Key decisions you made
- Any questions or concerns for review

Checking spawned agent results

After spawning, you can check progress:

# List all worktrees and their status
agent-cli dev status

# Read an agent's report
agent-cli dev run <branch-name> cat .claude/REPORT.md

# Open the worktree in your editor
agent-cli dev editor <branch-name>

Same-branch multi-agent workflow

Use this when several agents should inspect or validate the same code at once without separate worktrees.

# Create the worktree once. This step only prepares the shared workspace.
agent-cli dev new review-auth --from HEAD

# Then launch the actual agents with prompts. Give each agent its own tmux session.
agent-cli dev agent review-auth \
  --tmux-session review-auth-security-20260402-1530 \
  --prompt-file .claude/review-security.md
agent-cli dev agent review-auth \
  --tmux-session review-auth-performance-20260402-1530 \
  --prompt-file .claude/review-performance.md
agent-cli dev agent review-auth \
  --tmux-session review-auth-tests-20260402-1530 \
  --prompt-file .claude/review-tests.md

Key rules for same-worktree launches:

• Use dev agent, not dev new, after the worktree already exists
• Use dev agent --agent <agent> to select a specific agent for an existing worktree; --with-agent remains a deprecated alias on this subcommand
• For autonomous agents, prefer --tmux-session <unique-name> and use a different session name for each launch
• Reserve bare -m tmux for human-driven grouped windows or when the user explicitly wants agents to share one tmux session
• Outside tmux, bare -m tmux joins the deterministic repo-scoped tmux session
• Inside tmux, bare -m tmux opens a new window in the current session unless you pass --tmux-session <name>, which reuses or creates a specific tmux session and also implies -m tmux
• tmux session names cannot contain . or :
• Ask each agent to write to a unique report path such as .claude/REPORT-security-<run-id>.md or .claude/REPORT-tests-<run-id>.md
• If you rerun the same prompt repeatedly, include a timestamp or other run id in the report filename so later runs do not overwrite earlier ones
• Each agent launch gets its own unique task file in .claude/ (e.g., TASK-{timestamp}-{hex}.md), so parallel launches do not overwrite each other

Prompt guidance for shared worktrees

When multiple agents share a worktree, explicitly assign both a focus area and a unique report file. If you rerun the same review prompt often, prefer a timestamped filename such as .claude/REPORT-security-20260319-153045-123.md.

Prompt pattern:

Review the auth module for security issues only.

When complete, write findings to .claude/REPORT-security-20260319-153045-123.md including:
- Summary
- Issues found with file/line references
- Suggested fixes

Headless/scripted orchestration

For non-interactive contexts (scripts, cron jobs, other assistants), combine --prompt-file with --tmux-session:

agent-cli dev new validation-a --from HEAD --agent codex \
  --tmux-session validation-a-20260402-1530 \
  --prompt-file .claude/validation-a.md

This works without an attached terminal. For autonomous or scripted launches, prefer a unique --tmux-session per agent so separate runs do not trample each other by sharing one tmux session. Use bare -m tmux only when you explicitly want shared tmux grouping. Launches may also run pre-launch preparation by default; use --no-hooks only when you explicitly need to bypass that behavior.

Cleanup behavior

• dev rm and dev clean also clean up tmux windows that agent-cli tagged for the worktree
• If git worktree removal succeeds but tmux cleanup is partial, the command warns and still removes the worktree
• Session isolation prevents agents from interfering with each other's tmux windows during execution, but cleanup still applies to all tagged windows for the worktree

Example: Multi-feature implementation

If asked to implement auth, payments, and notifications:

# Spawn three parallel agents
agent-cli dev new auth-feature --prompt "Implement JWT authentication..."
agent-cli dev new payment-integration --prompt "Add Stripe payment processing..."
agent-cli dev new email-notifications --prompt "Implement email notification system..."

Each agent works independently in its own branch. Results can be reviewed and merged separately.

Key options

@examples.md