Claude Code Subagents in Practice — Working in Parallel

What subagents are

Claude Code can delegate a task to a subagent with its own context. Instead of the main session reading every file itself, you hand a broad task like exploration to a subagent and get back only the conclusion. That keeps the main context from being flooded with file dumps.

When they help

• Broad exploration: sweeping many directories and naming conventions to learn "where things live."
• Independent parallel work: running several non-dependent tasks at once.
• Adversarial verification: asking another agent to try to refute your result.

A practical pattern: explore → implement → review

1. Use an exploration subagent to map the relevant code and conventions.
2. The main session implements based on that conclusion.
3. A separate review agent independently checks the diff.

The key is "return only the conclusion." Have each agent make one decision at a time and leave the raw files in the subagent's context, so the main session stays light.

Beyond a single agent

Subagents are powerful, but they are ultimately coordinated sequentially inside one session. To run many agents physically at once and watch each one's progress at a glance, you need a separate orchestration layer. Marblo is built to run heterogeneous agents — including Claude — simultaneously on a kanban board, tracking each agent's state in real time.