multi-agent-patterns

This skill provides a robust framework for designing and implementing multi-agent architectures, specifically engineered to address the inherent context window bottlenecks of large language models. Rather than relying on single-agent setups, users can deploy this skill to distribute complex tasks across multiple specialized agents. It focuses on the core principles of context isolation, preventing the degradation of information quality known as the telephone game problem, and optimizing token economics in production-grade systems. This skill is essential for engineers, AI architects, and developers building scalable agentic systems that require high-fidelity coordination and parallel execution.

• Implementation of standard architectural patterns including Supervisor/Orchestrator for centralized control, Peer-to-peer/Swarm for flexible decentralized exploration, and Hierarchical structures for layered abstraction.

• Context management strategies to mitigate attention scarcity, context poisoning, and lost-in-the-middle degradation effects by partitioning information across lean, specialized agent nodes.

• Coordination protocol design for robust agent handoffs, consensus mechanisms that resist sycophancy, and failure handling to prevent cascading errors across the agent swarm.

• Token budget optimization techniques, comparing baseline single-agent costs against multi-agent systems to ensure performance gains justify the approximately 15x token multiplier overhead.

• Activate this skill when encountering task complexity that exceeds a single model's attention capacity or when tasks naturally decompose into parallelizable subtasks.

• Utilize this in workflows where specialized tool sets are required for different stages of the process, preventing a monolithic agent from carrying unnecessary context bloat.

• Ideal for production systems where human-in-the-loop oversight is required via a supervisor agent, or where highly autonomous research and exploration agents require peer handoff capabilities.

• When implementing, prioritize model selection alongside architecture design, as superior reasoning models often provide better performance returns than simply increasing context windows or total agent count.

• Be aware that supervisor-heavy designs may introduce latency; consider implementing a forward_message tool to allow direct feedback paths and improve fidelity.