agent-orchestration-patterns
Installation
SKILL.md
Agent Orchestration Patterns
Overview
Single agents hit hard limits: finite context windows, sequential throughput, and single-model reliability. Multi-agent systems unlock parallel execution, specialization, and fault isolation — but introduce new failure modes: coordination overhead, context duplication, inconsistent shared state, and cascading failures.
Use this guide when designing or reviewing systems where multiple AI agents coordinate, share work, or communicate results — whether agents are LLM-backed workers, rule-based processors, or hybrid pipelines. Applies to any system with a planner or orchestrator dispatching work to sub-agents.
Quick Reference
| Pattern | Topology | When to Use | Key Risk |
|---|---|---|---|
| Hierarchical Orchestrator | Orchestrator → N workers | Large tasks with clear decomposition | Orchestrator bottleneck; single point of failure |
| Flat Peer-to-Peer | Agents communicate directly | Negotiation, consensus, emergent behavior | Message explosion; hard to debug |
| Pipeline (Sequential) | A → B → C | Strict ordering, each stage transforms output | No parallelism; one failure halts all |
| Fan-Out / Fan-In | 1 → N parallel → 1 aggregator | Independent sub-tasks, time-sensitive results | Partial failures; aggregation complexity |
| Hybrid | Orchestrator + pipelines + peer links | Real production systems | All of the above |
| Supervisor Pattern | Monitor + restart agents | Long-running agents with known failure modes | Infinite restart loops |