Scaling Workflows

Part 4: Scaling Workflows

4.1 /batch — Codebase-Wide Migrations

/batch migrate all API routes from Express to Hono
/batch replace all uses of lodash with native equivalents
/batch add type annotations to all untyped function parameters
/batch standardize all API error responses

How it works:

1. Research and Plan: /batch explores affected files, decomposes into 5–30 independent units
2. Spawn Workers: One agent per unit in an isolated git worktree
3. Each worker: implement → run /simplify → run tests → commit → open PR
4. Merge independently — failure in one unit doesn't affect others

Prerequisites: Git repository required (uses worktrees for isolation). Add .claude/worktrees/ to .gitignore.

4.2 Agent Teams — Coordinated Parallel Work

Enable:

// In settings.json
"env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }

Usage:

Create an agent team for the payment integration:
- API agent: implement Stripe webhook handlers in services/payments/
- Frontend agent: build the checkout flow in apps/web/checkout/
- Test agent: write integration tests covering the full payment flow
Use Sonnet for each teammate to manage costs.

Display modes:

• In-process (default in non-tmux): Shift+Up/Down to switch between teammates
• Split panes (in tmux/iTerm2): Each teammate in its own pane
• Auto: Detects tmux → split panes, otherwise in-process

Team hooks:

• TeammateIdle: Fires when a teammate is about to go idle. Exit 2 to reassign work.
• TaskCompleted: Fires when a task is marked complete. Exit 2 to enforce quality gates.

Best practices:

• 3–5 teammates is the sweet spot
• Enable delegate mode to prevent the lead from doing implementation
• Assign clear file ownership per agent
• Token costs scale linearly with team size
• Use subagents for quick focused work; Agent Teams for coordination-heavy work

4.3 /loop — Recurring Monitoring

/loop 5m check the deploy status and report any errors
/loop 10m run the test suite and summarize failures
/loop 30m scan for new TODO comments added since last check

Session-level cron. Stays active while the session is open. Disable with CLAUDE_CODE_DISABLE_CRON.

4.4 /simplify — Post-Implementation Review

Run after finishing any feature, before opening a PR:

/simplify
/simplify focus on memory efficiency
/simplify focus on error handling

Spawns 3 parallel agents:

1. Code Reuse Agent: Scans for duplicated patterns, checks for existing utilities
2. Code Quality Agent: Looks for dead code, unclear logic, style issues
3. Efficiency Agent: Profiles for unnecessary allocations, redundant loops, batch opportunities

Findings are aggregated and fixes applied automatically. You review the diff.