
Multi-Agent Systems
Anthropic’s compiler experiment exposes the real work of agent handoffs
The 16-agent compiler prototype and Claude Research used different coordination mechanisms because they had different work to finish.
Coordination, delegation, and the costs of collaboration.
Anthropic’s compiler experiment exposes the real work of agent handoffs ↗

The 16-agent compiler prototype and Claude Research used different coordination mechanisms because they had different work to finish.

A compiler reference and a deliberately isolated permission checker illustrate why multiple agreeing models are not enough.

Anthropic’s April architecture separates durable sessions, execution, and the changing logic that directs long-running work.