git diff for Multi-Agent AI

Multi-agent discovery is not simply a matter of running eight coding agents at once. Shangyin Tan explains why long-horizon optimization needs diversity, and how his team avoids agents collapsing onto the same solution. For kernel optimization, agents work toward the same goal in separate sessions, with no direct point-to-point communication. Instead, each can inspect the others’ commits and scores. Restricting access to diffs rather than full files turns a code change into a transferable idea, helping weaker agents incorporate an algorithm without copying the entire implementation. The paper formalizes this as asynchronous, token-free communication, combined with periodic forgetting so the team restarts from its best candidate before accumulated history makes exploration conservative. Connor and Shangyin also discuss collaborative scratchpads, orchestrators, and why the right communication pattern depends on the problem. Clip taken from Weaviate Podcast #149: Scaling Discovery with Shangyin Tan -    • Scaling Discovery with Shangyin Tan - Weav...  .