Multi-Agent Orchestration

Decompose the task using horizontal (parallel), vertical (pipeline), or hierarchical (tree) patterns based on subtask dependencies. Delegate each subtask with a prompt specifying context, scope, expected output, and constraints, then synthesize results using merge, deduplication, or validation strategies.

Use a single agent for simple edits and capable-model multi-file refactors. Use multi-agent orchestration for research-plus-implementation work, cross-domain tasks that overload one context window, and independent work that can run in parallel.

Include four elements: context (project and relevant files), scope (what to do and not do), expected output format, and success criteria. Vague prompts like 'look for problems' produce unreliable results; specific prompts with JSON output schemas work best.

LangGraph suits state-machine workflows with cycles, AutoGen handles conversational research patterns, CrewAI fits role-based business processes, and VS Code agents integrate directly with IDE code tasks. Semantic Kernel serves enterprise .NET environments.

Common failures include over-orchestrating simple tasks, providing insufficient context to agents, lacking a synthesis strategy for outputs, and creating circular dependencies between agents. Start with a single capable model and add agents only when complexity demands it.