multi-agent-patterns

Multi-agent coordination distributes work across multiple language-model instances using supervisor, swarm, or hierarchical patterns to overcome single-agent context limits. It enables parallelism and specialization through task decomposition and explicit handoffs.

The supervisor or orchestrator pattern provides centralized control by routing tasks to specialized agents and synthesizing their results. It is ideal for workflows requiring aggregated outputs, such as routing tasks to a researcher, analyzer, and writer.

Context isolation partitions context and manages state across distributed AI agents by enforcing explicit handoffs, consensus mechanisms, and state transfer. This design principle prevents interference during parallel execution.

Robust multi-agent orchestration handles failure modes through explicit coordination protocols and consensus mechanisms. The provided references offer practical guidance on achieving convergence and reliable failure handling across supervisor, swarm, and hierarchical architectures.

Use a peer-to-peer swarm pattern for flexible handoffs and emergent problem solving, while hierarchical patterns suit layered strategy, planning, and execution. Choose based on whether your workflow needs decentralized flexibility or structured top-down control.

Yes, the Skill includes scripts and references to implement multi-agent coordination primitives like handoffs, state transfer, and messaging. It helps design workflows that aggregate results into final outputs across various domains.