Jev 作者发布 Jev 原生编码 agent 十步蓝图

sunick @Serantych原推发布于 2026.09.23

在 X 上打开原推

Jev Creator Diogo Almeida just released blueprint for building Jev-native coding agents 10-step blueprint for making coding agents cheaper, safer and easier to control step 1 → split the stack LLM writes and reasons, Jev handles bounded decisions, code executes deterministic actions step 2 → create a decision registry list every place your agent currently asks an LLM to choose, rank, approve, route or judge something step 3 → replace text decisions with typed outputs turn “what should I do next?” into Choice, Score or probability outputs your code can actually consume step 4 → build a live state snapshot feed Jev the current task, files touched, tool results, errors, permissions and previous actions instead of a giant conversational prompt step 5 → add confidence gates high confidence executes, medium confidence asks for review, low confidence blocks or falls back step 6 → route context before routing models decide what the next model actually needs to see before paying for a larger model step 7 → route tools dynamically load only the relevant tools and schemas for the current step instead of exposing the whole toolbox step 8 → make expensive models the exception use Jev to decide when a frontier model is really needed and when a smaller model or deterministic code is enough step 9 → record every decision store state → Jev output → confidence → action → result so bad decisions can be replayed and evaluated later step 10 → build an eval loop around outcomes measure whether Jev’s confidence matches reality, then tune thresholds instead of rewriting prompts the result: the expensive LLM stops making every tiny decision in the loop and becomes the specialist used only when reasoning or generation is actually needed

这条是引用着另一条写的：「Jev 作者 Diogo Almeida 谈自动化时代」