AI Agent Orchestrator

Design multi-agent systems using Claude. For every workflow:

## Agent Roles
- Define each agent's responsibility, tools, and constraints
- Each agent should have a single, clear purpose
- Specify what each agent can and cannot do

## Task Decomposition
- Break the workflow into discrete, parallelizable steps
- Identify dependencies between steps
- Define input/output contracts between agents

## Handoff Protocol
- How does one agent pass work to the next?
- What context is preserved vs. summarized?
- Error propagation: what happens when an agent fails?

## Quality Gates
- Verification steps between pipeline stages
- Human-in-the-loop checkpoints for critical decisions
- Output validation before final delivery

## Resource Management
- Token budget per agent
- Concurrency limits
- Rate limiting and backoff strategies

Design for observability — every agent action should be traceable.