拿到 Jev 权限后列出的 8 个 agent 构建用法

Layton Gott @Layton_Gott原推发布于 2026.09.17

在 X 上打开原推

One of the researchers behind ChatGPT just launched a model you can't chat with. It's called Jev. I also just got access and this is not what slowing the frontier looks like lol. It returns choices, scores, and probabilities instead of writing responses. Here are 8 things I'd build with it: For AI coders / agent builders: 1. A cheaper first-pass judge Give it the original task, the answer and the source material. Check instruction-following and whether the sources actually support the answer. Send questionable results for a closer review. 2. Retry or ask for help Give it the current step and the tool result. Check whether the step succeeded and whether information is missing. Your code uses those answers to retry, continue or ask a human. 3. Find the agent runs worth reviewing Give it the instructions, conversation and tool calls. Flag possible permission violations and unfinished tasks. Put those runs at the top of your review queue. 4. Stop dumping every search result into your prompt Give it the question and retrieved passages. Score their relevance. Pass the strongest matches to your main model. For founders / dev teams: 5. Prioritize inbound leads Give it the lead's message, company details and your customer criteria. Score fit and buying intent separately. Someone can fit your market without being ready to buy. 6. Route support tickets Give it the message and your available support queues. Return the department, urgency and whether a refund is being requested. Send uncertain cases to a person. 7. Make customer feedback countable Give it reviews or support messages and a fixed set of labels. Bug, feature request, pricing complaint, cancellation reason. Then count the recurring issues instead of asking for another vague summary. 8. Check replies before they reach customers Give it the draft, the customer's question and your policies. Check for unsupported promises, policy conflicts and questions left unanswered. Hold flagged drafts for review. The price: $0.042 per million input tokens. No output-token charge. And you can ask multiple independent questions about the same input in ONE request. My first test would be support routing: Take a small set of old, redacted tickets with known destinations. Have the agent build the classifier and show the wrong answers, response times and actual cost. Test it there before connecting it to a live inbox.

这条是引用着另一条写的：「ChatGPT 联合发明人发布新模型 Jev」