Morning Edition · Friday, November 6, 2026Published at 5:11 AM EST · New York

Octopus paper shares one KV cache across sub-agents and lifts BrowseComp-Plus accuracy to 49.6% on a 4-billion-parameter model

The same week, a second paper made reproducible inference cheaper under batching, and Anthropic cut the cost of Claude Code's background subagent status updates by about 99% after the tool had re-read whole transcripts every 30 seconds.

Multi-agent systems pay for context more than once. Octopus, a paper by Tianming Sha submitted November 3, identifies three costs in current frameworks: they copy prefix KV caches into every branch, they pass findings between agents as text…