| name | parallel-agents |
| description | Multi-agent orchestration patterns for Claude Code's native Agent tool. Use when coordinating multiple specialized agents for comprehensive reviews, cross-domain feature implementation, or parallel analysis from multiple perspectives. Adapted to this workspace's 42 actual agents. |
| allowed-tools | Read, Glob, Grep |
| metadata | {"triggers":"comprehensive review, multi-perspective analysis, parallel agents, orchestrate agents, coordinate agents, run multiple agents, full audit, cross-domain implementation","related-skills":"subagent-driven-development, multi-agent-patterns, multi-agent-brainstorming, architecture-design","domain":"workflow","role":"architect","scope":"design","output-format":"report"} |
| last-reviewed | 2026-03-15 |
Iron Law
USE THIS WORKSPACE'S ACTUAL AGENTS — never reference placeholder agents (security-auditor, backend-specialist) that don't exist here; always dispatch from the 42 agents in .claude/agents/
Explanation: Dispatching a non-existent agent silently fails. Every agent name in this skill maps to a real .md file in .claude/agents/.
Parallel Agents Orchestration
Orchestration guide for coordinating multiple specialized agents through Claude Code's native Agent tool. Keeps all orchestration within Claude's control using the built-in Agent tool and TeamCreate.
When to Use Orchestration
Good for:
- Complex tasks requiring multiple expertise domains (security + code quality + tests)
- Comprehensive reviews (architecture + security + accessibility)
- Feature implementation needing backend + frontend + database work
- Full pre-deploy audits across all stack layers
Not for:
- Simple, single-domain tasks (use the relevant specialist agent directly)
- Quick fixes or small changes within one file
- Tasks where one agent suffices
Decision rule: If dispatching more than 1 agent, use this skill for orchestration patterns and synthesis protocol.
Workspace Agent Catalog
Load references/workspace-agent-catalog.md for full descriptions. Quick reference:
| Agent | Domain | Use For |
|---|
code-reviewer | General | Quality, security, maintainability for any language |
security-reviewer | Security | OWASP Top 10, input validation, secret scanning |
threat-modeling-expert | Security | STRIDE analysis, attack trees, security architecture |
silent-failure-hunter | Error Handling | Swallowed exceptions, missing error states |
spring-reactive-reviewer | Java/Spring | Reactive correctness, Resilience4j, R2DBC |
nestjs-reviewer | NestJS | Module correctness, JWT, Prisma patterns |
riverpod-reviewer | Flutter | Provider types, ref usage, AsyncValue handling |
agentic-ai-reviewer | LangGraph | Graph correctness, guardrails, cost, production readiness |
postgresql-database-reviewer | PostgreSQL | Query optimization, schema, security, performance |
pgvector-schema-reviewer | pgvector | Vector schema, index alignment, embedding model |
weaviate-schema-reviewer | Weaviate | Collection definition, vectorizer, multi-tenancy |
rag-pipeline-reviewer | RAG | Chunking, model pinning, reranking, error handling |
accessibility-auditor | Accessibility | WCAG 2.1, Semantics, focus management |
ui-standards-expert | UI/UX | Design tokens, theming, touch targets |
flutter-security-expert | Flutter | Secure storage, certificate pinning, GDPR |
dedup-code-agent | Tech Debt | Duplication, unused code, dependency bloat |
plan-challenger | Planning | Adversarial plan review (5 attack dimensions) |
Orchestration Patterns
Load references/orchestration-patterns.md for full pattern details.
Pattern 1: Comprehensive Feature Review
For any feature touching security-sensitive code, database, or multiple services:
Step 1: code-reviewer → overall quality gate
Step 2: [stack-specific reviewer] → framework-specific patterns
Step 3: security-reviewer → OWASP + auth + input validation
Step 4: silent-failure-hunter → error handling completeness
Step 5: pr-test-analyzer → behavioral test coverage
Step 6: Synthesize all findings
Pattern 2: Pre-Deploy Audit
Before merging any significant feature branch:
Parallel:
security-reviewer → security findings
[stack reviewer] → framework correctness
pr-test-analyzer → test coverage gaps
Sequential:
output-evaluator → APPROVE / NEEDS_REVIEW / REJECT verdict
Pattern 3: Full Flutter Feature
riverpod-reviewer → state management
flutter-security-expert → secure storage, GDPR
accessibility-auditor → WCAG 2.1
ui-standards-expert → design tokens, touch targets
silent-failure-hunter → error handling in async code
Pattern 4: Agentic AI System Review
agentic-ai-reviewer → graph correctness, guardrails
security-reviewer → prompt injection, input validation
rag-pipeline-reviewer → chunking, model pinning (if RAG used)
silent-failure-hunter → tool error handling
Synthesis Protocol
Load references/synthesis-protocol.md for the full template.
After all agents complete:
## Orchestration Synthesis
### Task Summary
[What was accomplished]
### Agent Contributions
| Agent | Key Finding |
|-------|------------|
| [agent] | [finding] |
### Consolidated Recommendations (severity-ordered)
1. **CRITICAL** — [issue] — must fix before merge
2. **HIGH** — [issue] — should fix before merge
3. **MEDIUM** — [issue] — fix in follow-up PR
4. **LOW** — [issue] — note for tech debt backlog
### Verdict
[APPROVE | NEEDS_REVIEW | BLOCK] — based on finding severity
Constraints
MUST DO
- Only dispatch agents that exist in
.claude/agents/ (42 agents listed in catalog above)
- Pass explicit context to every agent — agents do NOT inherit CLAUDE.md or conversation history
- Produce a single synthesized report, not separate outputs per agent
- Include test coverage review (pr-test-analyzer) for any code-modifying orchestration
MUST NOT DO
- Dispatch agents with only "review this code" — always pass specific context and scope
- Run more agents than needed — match agent count to actual coverage gaps
- Skip synthesis — raw agent outputs are not a deliverable
Knowledge Reference
Agent orchestration, multi-agent coordination, parallel dispatch, reviewer agents, synthesis protocol, pre-deploy gate, Claude Code Agent tool, TeamCreate, context passing
