Stanford's Azalia Mirhoseini releases DeLM, a decentralized agent framework that improves SWE-bench accuracy by 10% while halving costs

more ppl are now trying out this approach of agents communicating with a shared state (vs talking to each other)

@yoheinakajima Glad to see more people playing with shared state architectures for agents.

Feels like this pattern is showing up independently in a lot of places now.

@StanfordAILab @VentureBeat @Mao_Yuzhen @Azaliamirh How do you make sure all these agents are secure against things like prompt injection?

https://venturebeat.com/orchestration/stanfords-delm-cuts-multi-agent-task-costs-50-without-a-central-orchestrator

@christinetyip Indeed!

@yoheinakajima Huh so deciding where to fork the agent for new subtask becomes more important?

@yoheinakajima shared state makes more sense than watching agents talk past each other

@yoheinakajima post swarm state could be something emulating different tiers of shared states (only communicating laterally or upwards) similar to the NOW window approach

@yoheinakajima yeah because debugging telephone-game between agents is hell. shared state and you can actually see what everyone knows

@StanfordAILab @VentureBeat @Mao_Yuzhen @Azaliamirh 👍🏼

@yoheinakajima yep, operational state + coordination layer

this is the way

we are betting on graph: https://github.com/ModernRelay/omnigraph

@yoheinakajima ok boss

@yoheinakajima shared state makes more sense anyway, talk between agents just adds overhead for no reason

@yoheinakajima Agreed. I think the real shift is from orchestration to shared infrastructure. Once progress is persistent and verified, parallel work becomes much more composable.

@StanfordAILab @VentureBeat @Mao_Yuzhen @Azaliamirh I'm wondering if this is basically mixture of experts at the multi-agent level now. Same pattern, different abstraction layer.

@yoheinakajima Same lesson distributed systems learned decades ago. Central coordination doesn't scale. Why would AI agents be different?

@yoheinakajima We'll probably see a new wave of distributed event streaming platforms because kafka can be a pain to deal with.