Swarm
Multi-agent pipelines — planner, coder, tester, and reviewer collaborating on a task with visual progress tracking.
@elyracode/swarm orchestrates structured, multi-stage workflows where each stage is a focused agent turn. Results flow from one stage to the next as context, so you get sharper outcomes than asking a single generalist to do everything.
Install
elyra install npm:@elyracode/swarm
Pipelines
Build (/swarm build <task>)
Full feature development pipeline:
plan → code → test → review → fix
- Plan — Analyzes the codebase, creates an implementation plan with file paths
- Code — Implements the plan and edits files
- Test — Writes and runs tests for the implementation
- Review — Reviews changes for correctness, security, and edge cases
- Fix — Addresses review findings
Review (/swarm review <target>)
Deep multi-pass code review:
analyze → correctness → security → tests → synthesize
- Analyze — Maps code structure and data flow
- Correctness — Reviews logic, error handling, edge cases
- Security — Checks for vulnerabilities, input validation, auth issues
- Tests — Evaluates test coverage and quality
- Synthesize — Prioritizes findings into an actionable list
Refactor (/swarm refactor <target>)
Structured refactoring:
analyze → plan → implement → verify
- Analyze — Understands current structure and identifies issues
- Plan — Creates a refactoring plan that preserves behavior
- Implement — Executes the refactoring
- Verify — Confirms behavior is preserved, no regressions
Commands
| Command | Description |
|---|---|
/swarm |
List available pipelines |
/swarm build <task> |
Run the build pipeline |
/swarm review <target> |
Run the review pipeline |
/swarm refactor <target> |
Run the refactor pipeline |
Examples
/swarm build add a notification system with email and in-app channels
/swarm review the authentication module
/swarm refactor the database query layer
Natural language
Just describe what you want and mention "swarm":
Use swarm to build a REST API for user profiles
Run a swarm review on the checkout flow
How it works
Each pipeline stage runs as a focused agent turn with specific instructions. The agent sees a progress indicator showing which stage is active, and results from each stage are passed to the next as context.
Read-only stages (analyze, review, synthesize) cannot edit files. Only implementation stages (code, fix, implement) can write.
Related
- Subagents — lower-level building blocks for delegating to focused child agents