Parallel Agents

What is Parallel Agents

Parallel agents are multiple agent sessions running at the same time, each working an independent piece of a larger task in its own context window. Instead of one agent handling five subtasks one after another, five agents each handle one, and the total wall-clock time drops toward the length of the longest single piece rather than the sum of all five.

The gain is specifically in wall-clock time, not in the total amount of work performed, five agents working five subtasks still do roughly five subtasks worth of work, just concurrently instead of sequentially. This only pays off when the pieces are genuinely independent, if one subtask needs the output of another, running them in parallel just means one of them waits or produces something that has to be redone.

Isolation is the other half of the picture: parallel agents editing the same repository need something like separate worktrees to avoid colliding on the same files. Without that isolation, running agents in parallel trades a slower single agent for a faster but riskier set of agents stepping on each other's changes.