3/3
At 200 tools that split was ~$0.03 (Jev + Kimi K3) vs ~$0.14 (Kimi K3) and ~$0.24 (GPT-6 Astra).
System 1 for the choice. System 2 for the heavy work.
Blog: https://t.co/F1wtshtWbz
Code + results: https://t.co/4N48BtBXbG
cc @typesafeai
1/3
Does your agent need to see ~100 tools every turn?
That burns tokens, context and money.
I swapped “which tool next?” to @typesafeai 's Jev. At 200 tools, Jev + Kimi was ~$0.03 vs ~$0.14 (Kimi K3) and ~$0.24 (GPT-6 Astra). Same tasks, same fakes.
🧵
2/3
Jev makes the decision for the agent on the next tool call.
The LLM never sees the full menu. It only fills arguments for the one tool Jev picked — or writes the note when Jev says finish.
2/2
Not a giant eval — only 3 tasks — but the routing split is the point: System 1 for the choice, System 2 for the work.
Blog: https://t.co/F1wtshtWbz
https://t.co/c1lRL4an5r
Walmart has a massive Apache Kafka deployment with 25K+ consumers across private and public cloud environments. In this post, we’ll look at the main challenges of a Kafka setup at this scale.