What problem does it solve? Wide tasks like audits across many files, sweeps for one class of fault, or findings that must be refuted before they are believed exceed what a single agent can do reliably. This Skill defines how to cut such work into parallel agent strands that run fast and come back trustworthy. ## Core Features & Use Cases - Pipeline over barrier: Uses pipeline() so each item runs its whole chain independently, reserving parallel() barriers only for stages that genuinely need all prior results at once. - Finder-refuter pattern: Every finding goes to an agent whose job is to prove it wrong, with structured verdict schemas and rejections reported with reasons. - Cost control and resumability: Counts strands and estimates tokens before a run, asks past the 2.5 million token line, and has strands save after every block so a stop costs a block, not a strand. - Use Case: Auditing every test file in a repository for one class of fault: scout the file list with grep, fan out one reader per file at low effort, then send each file's findings to a single high-effort reviewer that attacks them all. ## Quick Start Use the workflow skill to plan and run a multi-agent audit across these files, with a refuting agent checking every finding.
