skills.rest
jnyross
jnyrossCOMMUNITY
Explore all skills by this author →
GitHub

swarm

Fan out parallel cloud workers and aggregate their results into one report.

Updated Sep 2, 2026
Search Tags:#parallel agents#aggregation#task decomposition#swarm#fan out#model race#cloud workers
One-click install
npx skills add https://github.com/jnyross/pstack-muse --skill swarm-jnyross
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: swarm
Source: https://github.com/jnyross/pstack-muse/tree/main/skills/swarm
Command: npx skills add https://github.com/jnyross/pstack-muse --skill swarm-jnyross

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Coordinating many parallel AI workers on a large task is error-prone: slices get missed, race results go unranked, and raw worker output floods the conversation. This Skill provides a disciplined four-phase workflow to frame, launch, aggregate, and report on N parallel cloud workers. ## Core Features & Use Cases - Structured Fan-Out: Spawn N background cloud workers in one message, each with a standalone brief covering goal, scope, verification, and reporting format. - Coverage, Race, and Mixed Shapes: Partition work into slices, race N workers on identical briefs with a declared selection rule (first pass, rank all, or best-of), or mix both. - Consolidated Reporting: Aggregate terminal results into a compact table with evidenced one-line issues, explicit gaps, and dropout notes instead of raw worker dumps. - Use Case: You need a codebase audited across five modules. Launch five workers, one per module, each writing to its own worktree, then receive a single report with PASS/ISSUES/BLOCKED status per module. ## Quick Start Ask the assistant to swarm this task by fanning out parallel workers to cover each module and return one consolidated report.

Frequently Asked Questions about swarm

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run parallel AI workers on a large task?▼

Frame a done predicate, choose a shape (slices, race, or mixed), set N workers, and spawn them all in one message as background cloud agents. Each worker gets a standalone brief with goal, scope, verification, and reporting format, then results are aggregated into one report.

What is a model race with parallel agents?▼

A model race runs N workers on identical briefs using different models, with a selection rule declared up front: first pass, rank all, or best-of. Each arm's model is named before spawning, and the declared rule determines which result wins.

How do parallel workers avoid overwriting each other's changes?▼

Each worker that writes gets its own writable output: a separate worktree, branch, or a directory like /tmp/swarm-<slug>/worker-<n>/. This isolation is assigned during the framing phase before any worker launches.

What happens when a swarm worker drops out or fails?▼

The run proceeds with N-1 workers and the dropout is noted explicitly. The final report includes a gaps or dropouts section so missing coverage is visible rather than silently omitted.

When should I use local instead of cloud workers?▼

Use environment "local" only when a worker needs access to something on the user's computer. Cloud is the default for parallel fan-out, and workers needing a non-default pushed branch can receive a cloud_base_branch parameter.

Related Skills

Explore More Software Engineering Skills→
jeremybrasherjeremybrasher

swarm

Fan out N parallel cloud workers and aggregate results into one consolidated report.

Community
Intermediate
miles990miles990

verification-before-completion

Run verification commands and report results only when checks pass.

Community
Advanced
QunnnnQunnnn

dispatching-parallel-agents

Coordinate parallel delegation of independent tasks across specialized agents.

Community
Intermediate
OpenCovenOpenCoven
20

serving-llms-vllm

Serve LLMs with OpenAI-compatible endpoints, quantization, and tensor parallelism.

Official
Advanced