12 个用 Jev 替代提示词的 agent 模式
列出浏览器下一步动作、keep-or-drop 压缩、技能按需加载、引用审计、语义 CI 闸门等 12 种模式，Jev 只给类型化裁决加概率，流程仍由普通代码跑。
12 agent patterns where Jev beats a prompt (save this) Jev handles the semantic calls that if-statements cannot encode and that a chat model is too slow or too sloppy to own. You get a typed verdict plus a probability. Ordinary code still runs the workflow. 12 patterns that keep showing up: 1. Next-step in the browser > Turn the live DOM into one bounded action - click, type, or halt. Code executes only legal operation-target pairs. 2. Keep-or-drop compaction > Decide which events from a long trace survive. Survivors stay verbatim. Nothing gets rewritten as a summary. 3. Just-in-time skill load > Compare this user turn to the skill catalog. Inject only the instructions that apply, not the whole library. 4. Typed call compiler > Map a natural-language request to one function and fill each typed argument on its own. Code runs only after every slot clears. 5. Citation auditor > Check that the quoted passage exists and that nearby evidence actually carries the claim. Output: supported, thin, or contradicted. 6. Extract-then-verify > Run a cheap extractor first. Jev only sees questionable fields. Clean records stay on the fast path. 7. Trace labeler > Stamp a raw trajectory with queryable labels — progress, stall, loop. Skip the second LLM that writes a full review of every run. 8. Semantic CI gate > Replay a golden trace suite against a new agent build. Semantic checks can pass or block prompt, model, tool, and policy changes. 9. Meaning-first code search > Search a repo by what the code does, not the tokens it happens to contain. Rank candidate snippets and return the useful ones first. 10. Record identity referee > Two candidate records in. Verdict: merge, review, or keep separate. Candidate generation stays deterministic; Jev owns identity. 11. Evidence reranker > Embeddings retrieve a wide set. Jev reorders passages by whether they actually answer the question, then generation sees the good ones first. 12. Lesson promotion lock > Take a finished agent trace and judge whether its corrections contain a reusable lesson. Promote the generalizable ones. Discard task noise. The last pattern is already wired up in Beacon (open source). It records full sessions from Claude Code, Codex, Cursor, OpenCode and 20+ other harnesses. Jev then scores which corrections are reusable lessons so a fix found in one agent can land in the others.
这条是引用着另一条写的：x.com/merccante/status/2102478458727113139