Five Months In: Why Parallel Claude Code Beats Nested Subagents
Five Months In: Why Parallel Claude Code Beats Nested Subagents
After five months of trying to make Claude Code subagents work, I gave up on nesting entirely. The limitation is fundamental - subagents cannot spawn their own subagents, which means any task that needs more than one level of delegation hits a wall.
The Nesting Problem
The idea behind subagents is appealing: your main agent delegates subtasks to specialized child agents. But in practice:
- Depth limit of one - a subagent cannot create its own subagents, so complex tasks cannot decompose further
- Context isolation - subagents start with limited context and spend tokens re-discovering things the parent already knows
- Coordination overhead - the parent agent spends more time managing subagents than it would just doing the work
- Failure cascading - when a subagent fails, recovering at the parent level is messy
The Parallel Process Alternative
Instead, I run 5 separate Claude Code processes in parallel on the same repo. Each one gets a clear, scoped task:
- Agent 1: implement the API endpoint
- Agent 2: write the frontend component
- Agent 3: update the tests
- Agent 4: fix the database migration
- Agent 5: update the documentation
They all work on the same codebase simultaneously. There is no nesting, no parent-child relationship, no coordination protocol. Just 5 independent workers.
Making It Work
The key is task isolation. Each agent works on different files. When there is overlap, you handle merge conflicts manually - which happens less often than you would expect if you scope tasks well.
Use tmux or separate terminal tabs to monitor all five. Each agent has its own CLAUDE.md context tailored to its specific task.
The Results
What used to take 3-4 hours with sequential work now takes about 40 minutes. The parallelism is real, and the simplicity of independent processes beats the complexity of nested orchestration every time.
Fazm is an open source macOS AI agent. Open source on GitHub.