Claude Code Subagents: Spawning Specialists That Actually Help

The single-agent pattern breaks around feature four. The context window fills with old decisions, unrelated files, and noise. What I thought was agent fatigue was really my own prompting fatigue. Subagents fixed it for me.

The Core Shift

A subagent is just a fresh Claude instance with a narrow mandate and its own context. I use the orchestrator agent to read the plan, then spawn specialists to execute. The specialist does not know about the other tasks, and that is the point. Smaller scope means sharper output.

In my current workflow I spawn three classes of subagent: an explorer that reads and returns findings, an implementer that writes code against a spec, and a reviewer that checks the diff. Each one gets maybe 2000 tokens of instructions and nothing else.

A Real Spawn Pattern

# Orchestrator plans, subagent executes
task = {
    'role': 'implementer',
    'scope': 'Write POST /subscribers endpoint only',
    'files_allowed': ['app/routes/subscribers.py', 'tests/test_subscribers.py'],
    'return': 'diff + test results',
}
# Spawn with just this context. No history. No other tasks.

The explicit files_allowed list is what makes this work. When the subagent finishes, it returns a structured result to the orchestrator, and the orchestrator decides what to spawn next. You get parallelism almost for free.

Why It Beats One Big Agent

One big agent has to hold the entire plan in context. Five subagents each hold one slice. When one slice goes wrong, I re-spawn just that slice instead of unwinding the whole session. Debugging costs drop hard.

Read more at the Claude Code subagents guide. Start with two subagent roles and add a third only when you feel the pain.