Jev 创始人 12 页 PDF：编码 agent harness 十步蓝图

rari @0xwhrrari原推发布于 2026.09.23

在 X 上打开原推

Jev Founder, Diogo Amogo, just released 12-page PDF on building a Jev Harness for coding agents This is a 10-step blueprint on how to make your coding agents 220x faster and 444x cheaper: step 1 → stop treating a coding agent like a chatbot with shell access: the model proposes actions, while the harness controls state, tools, permissions, budgets and completion step 2 → specify the work before execution: every task needs a goal, scope, expected artifacts, allowed side effects, verification commands and a strict definition of done step 3 → treat context as a working set: load only the files, instructions, messages and tool outputs required for the current decision instead of dumping the entire repository into every turn step 4 → expose tools through policy: every action should pass through an allow / ask / deny gate based on its arguments, affected files and potential consequences step 5 → begin with one capable agent: add planners, implementers and reviewers only when the work can be separated into independent units with clear ownership step 6 → isolate parallel work: give every agent its own worktree, write set and integration contract so speed does not turn into duplicated edits and merge conflicts step 7 → require evidence before completion: a confident final message means nothing without the diff, test output, exit status, review findings and replayable verification commands step 8 → engineer memory and recovery: persist decisions, checkpoints, budgets and unresolved risks so long-running tasks can survive compaction, interruption and failed attempts step 9 → make the system observable: record context selection, model routing, tool calls, retries, costs and verification results before increasing autonomy step 10 → convert failures into infrastructure: every incident should produce a new policy, regression test, evaluation case, recovery path or tighter task contract Most courses teach you how to prompt a coding agent This 12-page PDF teaches you how to build one that can be inspected, resumed, evaluated and trusted in production Send Jev Engineering for Coding Agent to your Claude Code or Codex instance and use it as the blueprint for your next agent system ↓

这条是引用着另一条写的：「Jev 把路由打分验证做成决策层」