Agent loops make a full LLM call even for a simple yes/no decision — that's a lot of waste.
cv usk@cv_usk
Pattern↑ Jev AI decision making
Agent loops make a full LLM call even for a simple yes/no decision — that's a lot of waste. Here's a different way to handle those decisions.
Title: Building a Harness with Jev
URL: langchain.com/blog/building-a-harness-with
❓ What's new about Jev?
💡 It's a "System One model" from TypeSafe AI that doesn't generate text at all. You feed it a state and a set of questions, and it returns typed answers with probabilities.
❓ How is it different from a regular LLM?
💡 It's trained via reinforcement learning and specializes in three question types: Choice (pick from options), Score (rate on a scale), and Noul (yes/no). Bundling multiple questions into one call barely adds latency or cost.
❓ Where would you actually use this?
💡 Two big ones: model routing (send simple tasks to cheap fast models, hard ones to capable models) and guardrails (checking risk before letting an agent run something like a bash command).
❓ How much does it really help?
💡 On classification tasks it reports 200x faster inference and 400x lower cost, and it's already running in production for browser automation, live trading agents, and large-scale email classification.
#AIAgents #LangChain