swarm-speculative

You can compare multiple code refactoring strategies by running parallel agentic implementations in isolated worktrees and evaluating each approach via a judge agent to select the optimal solution. This handles scenarios where the optimal strategy is unclear by testing multiple viable implementations simultaneously.

Best-of-n parallelism for code generation spawns multiple agents to work on the same goal using different strategies simultaneously. You should use it when facing risky changes, unclear optimal strategies, or when comparing multiple viable implementations for quality and performance.

You can evaluate competing code implementations automatically by using a designated judge agent that assesses each parallel implementation based on predefined criteria. The judge agent performs comparative analysis to identify and report the most effective approach.

Parallel agent execution requires isolated worktree execution for each implementation to prevent conflicts. This agentic coordination ensures task delegation remains separated while allowing a judge agent to safely perform comparative analysis on each approach.

You should avoid using parallel agents for code refactoring when the optimal strategy is already clear, changes are low risk, or only one viable implementation exists. This approach requires agentic coordination overhead best reserved for uncertain or complex architectural decisions.