Arc provides a rich set of workflow patterns for orchestrating multi-agent systems. Each pattern addresses specific collaboration scenarios and offers unique advantages for different use cases.
Available Patterns
Sequential Workflow
Execute agents one after another, where each agent builds upon the previous agent's output.
Best for: Linear processes, step-by-step transformations, pipeline processing
Concurrent Workflow
Run multiple agents in parallel and aggregate their results.
Best for: Independent tasks, parallel processing, consensus building
Mixture of Agents
Combine multiple agent perspectives through iterative refinement layers.
Best for: Quality improvement, multi-perspective analysis, iterative refinement
Forest Swarm
Organize agents into a tree structure with specialized sub-teams.
Best for: Hierarchical decomposition, complex problem breakdown, divide-and-conquer
Hierarchical Swarm
Structured hierarchy with explicit manager-worker relationships.
Best for: Enterprise workflows, clear authority structures, delegation patterns
Heavy Swarm
Dense collaboration network where all agents can interact with each other.
Best for: Brainstorming, consensus building, peer review systems
Group Chat
Turn-based conversation system with dynamic agent selection.
Best for: Interactive discussions, role-playing, conversational workflows
Swarm Router
Intelligent routing system that dynamically selects the best agent for each task.
Best for: Task classification, expert selection, adaptive workflows
Pattern Comparison
| Pattern | Complexity | Parallelism | Use Case |
|---|---|---|---|
| Sequential | Low | None | Linear pipelines |
| Concurrent | Low | High | Independent tasks |
| Mixture of Agents | Medium | High | Quality refinement |
| Forest Swarm | Medium | Partial | Hierarchical breakdown |
| Hierarchical Swarm | Medium | Partial | Enterprise workflows |
| Heavy Swarm | High | Full | Dense collaboration |
| Group Chat | Medium | None | Interactive discussions |
| Swarm Router | Low | None | Dynamic routing |
Choosing a Pattern
Consider these factors when selecting a workflow pattern:
- Task Dependencies: Are tasks sequential or independent?
- Coordination Needs: How much do agents need to communicate?
- Scalability: How many agents will participate?
- Output Quality: Is consensus or refinement needed?
- Latency: Can you afford parallel execution overhead?
Common Use Cases
Content Generation
- Sequential: Draft → Edit → Proofread
- Concurrent: Multiple writers → Aggregator
- Mixture: Multiple drafts → Refinement layers
Research and Analysis
- Forest Swarm: Topic breakdown → Specialized research
- Heavy Swarm: Peer review and consensus
- Swarm Router: Query routing to experts
Decision Making
- Hierarchical Swarm: Manager delegation
- Group Chat: Stakeholder discussion
- Concurrent: Independent evaluations → Vote
Code Development
- Sequential: Design → Implement → Test → Deploy
- Forest Swarm: Feature breakdown → Parallel development
- Group Chat: Code review discussions