A seven-step guide to giving agents a decision layer, plus a Jev vs OpenAI Decisions test
The guide covers finding decision points in 20 traces, writing each answer as a condition plus a none option, setting confidence floors per action, and comparing Jev with OpenAI Decisions in Every's test, where they scored 76 and 73 correct steps out of 78.
OpenAI just entered the decision-model race that Jev made popular how to give your agent a decision layer that picks every next step, in this guide: step 1 -> find the decisions: mark every step in 20 traces where your agent picks from a short list step 2 -> write each answer as a condition a stranger could apply, plus a none option for the gaps step 3 -> set floors per action: 0.6 to act on a read, 0.85 before money moves without a confirm step 4 -> on Jev, ask every question in one call, then let code ignore what doesn't apply step 5 -> send everything under the floor to code, a bigger model or a person step 6 -> replay a few hundred real decisions through each engine - Every's two tests split the win step 7 -> pin the model version - a floor tuned on one model is a guess on the next 76 vs 73 correct steps of 78, OpenAI's Decisions against Jev, in Every's first test the result: each call comes back in under a second, and the big model only sees the hard cases a must-have for anyone building agents, read the full guide below
This post quotes another post: x.com/k4zxbt/status/2105378691614175362