sadd:multi-agent-patterns

Multi-agent patterns distribute complex tasks across specialized agents to overcome single-agent context limits. You need them for large research, code reviews, or planning workflows that require decomposing tasks into parallel subtasks.

To coordinate subagents using file-based memory, you isolate contexts and use the file system as the primary communication mechanism. This enables scalable collaboration by allowing agents to exchange structured outputs without shared context limits.

The best way to design a hierarchical pipeline is to use an orchestrator that partitions work into subtasks and delegates them to specialized subagents. This coordination pattern improves throughput by isolating contexts and managing dependencies explicitly.

Yes, you can use centralized coordination patterns for large research tasks in Claude Code. A supervisor pattern distributes the workload across subagents, allowing them to explore specialized areas in parallel while returning structured outputs to the orchestrator.

Single-agent reasoning fails on complex code reviews due to limited planning horizons and context constraints. A peer-to-peer swarm pattern helps by partitioning the codebase among specialized agents that review distinct sections in parallel.

The limitation of context isolation is that subagents cannot implicitly share state, requiring explicit file-based memory for inter-agent communication. This adds coordination overhead but prevents individual agent context limits from bottlenecking the workflow.