Multi-Agent Architectures
Specialized roles under an orchestrator mirror an investment committee, at a real cost in tokens, latency, and failure modes.
Multi-agent systems split work across specialized roles, where a planner decomposes the task, researchers gather evidence, and a critic checks the draft, coordinated through an orchestrator or through direct handoffs. The pattern maps naturally onto a desk investment committee: bull and bear analysts argue a thesis, a risk officer challenges assumptions, and a chair decides. The trade-offs are concrete: more agents mean more tokens, latency, and failure surface, and orchestration adds its own bugs in loops, dropped context, and agents talking past each other. Start single-agent, add roles only when specialization demonstrably beats one well-tooled loop, and keep humans approving anything that moves money.