A paper: Language Model Teams as Distributed Systems
- A paper: Language Model Teams as Distributed Systems
- TL;DR: treat your agent orchestrator like a job scheduler in a distributed system.
- Topological sort your DAG of tasks.
- Assign via round-robin respecting dependencies.
- Add a straggler timeout that triggers reallocation.
- Minimize inter-agent messaging to only what's necessary for shared state consistency.
- Like humans, communication blows up at n^2 rate.
- Centralized coordination works best.
- Swarms are best for high-beta tasks.
- That is, a high rate of failure, also a high rate of great results.
- Swarms use significantly more tokens; if it's not parallelizable, it doesn't lead to any improvements.