# AI Agent Orchestration
AI agent orchestration is the coordination of multiple [[AI Agents]] working together on a task. Rather than a single agent handling everything, orchestration splits work across specialized agents, each with its own context, tools, and instructions, then combines their outputs.
The [[Receptionist AI Design Pattern]] is a common orchestration pattern: a routing agent receives user input and delegates to specialized agents. [[Claude Code]] uses a simpler form with subagents: the main agent spawns child agents for parallel tasks (research, testing, exploration), each inheriting a subset of context.
Orchestration makes [[Separation of Concerns]] concrete in AI systems. Each agent operates within its own [[Context Window]] with its own [[Token Budget]], loaded via [[Prompt Lazy Loading AI Design Pattern (PLL)]]. This avoids [[Context Bloat]] from cramming every capability into a single prompt, and allows different agents to use different models optimized for their specific task.
The challenges are coordination overhead, context sharing between agents (what does agent B need to know about agent A's output?), and error handling across agent boundaries. These are analogous to microservice orchestration challenges in distributed systems.
## Orchestration Patterns
**Predefined teams**: Agents are configured in advance with a shared goal, execution mode (parallel or pipeline), and handoff protocol. Good for recurring workflows.
**Panels**: Multiple agents evaluate the same content independently, producing a combined scorecard. Good for quality gates and multi-angle review.
**Emergent chaining**: Agents discover work for other agents during execution and suggest handoffs via a standardized protocol. The orchestrator validates (no loops, depth limits) and chains automatically. Good for organic collaboration where the full workflow isn't known upfront. Requires safety rules: max chain depth, no duplicate agents, no circular patterns.
**State persistence**: Agents that run multi-step flows across sessions can persist in-progress state (current phase, collected data, next steps) in a lightweight state file, enabling resumption without re-asking questions or losing progress.
## References
-
## Related
- [[AI Agents]]
- [[Receptionist AI Design Pattern]]
- [[AI Agent Harness]]
- [[AI Agent Skills]]
- [[Claude Code]]
- [[Separation of Concerns]]
- [[Context Engineering]]
- [[Token Budget]]
- [[Prompt Lazy Loading AI Design Pattern (PLL)]]
- [[Context Bloat]]
- [[Composition over Inheritance]]
- [[Loose Coupling]]