The Serial Workflow Trap
TRIGGER
Engineers using AI coding agents spend significant time waiting—the agent processes one task while the human sits idle, creating a bottleneck where expensive human attention is blocked by AI compute time rather than being utilized on other work.
APPROACH
Anthropic engineers maintain 3-4 git worktrees (or full repo checkouts) in separate folders, each with its own terminal tab running Claude on independent tasks. Workflow: (1) Create worktrees via `git worktree add ../project-feature-a feature-a`, (2) Launch Claude in each: `cd ../project-feature-a && claude`, (3) Cycle through tabs to check progress and approve/deny permission requests. Best practices: use consistent naming conventions, maintain one terminal tab per worktree, set up notifications for when Claude needs attention, use separate IDE windows for different worktrees, clean up when finished with `git worktree remove`.
PATTERN
“Expensive human attention blocked while the agent computes—the "serial workflow trap" where you sit idle waiting for one task to finish. The bottleneck is human review, not agent compute. Run 3-4 agents in parallel worktrees; cycle through tabs approving and course-correcting. Maximize human utilization, not agent utilization.”
✓ WORKS WHEN
- Tasks are independent and don't modify overlapping files
- Human review/approval is the bottleneck, not agent compute
- You can context-switch between 3-4 different tasks without confusion
- Git worktrees are supported (tasks on different branches)
✗ FAILS WHEN
- Tasks have dependencies and must be sequenced
- Cognitive overhead of tracking multiple concurrent tasks exceeds time saved
- Tasks modify shared files, creating merge conflicts
- Single long-running task requires continuous human attention (debugging session)