📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
hn · Ask HN · 43 分 ← 返回列表

Ask HN: How to get the most intelligence out of your agents?

原文链接 ↗ ·thechairman12
I often find chatgpt.com has greater "intelligence" than codex. I've found this to be true for the past year or two now, and I was wondering if anyone has found a better solution than copying context to and from chatgpt.com. The "Pro" level (or whatever the max level is at the time) consistently has better insights than the agent, but immensely limited tool calling compared to the agent. Marrying the two of these together unlocks a lot of potential, but I really just want this intelligence to li

五维评分

付费意愿
35
需求真实度
42
个人可落地
72
竞争空间
28
新鲜度
58

AI 分析

痛点:高价的高级推理模型(chatgpt.com Pro)洞察力强但工具调用能力弱,而本地 agent(codex)工具调用强但推理弱,用户被迫手动来回复制上下文。

目标用户:重度使用 AI 编码 agent 的个人开发者与小团队工程师

证据:I often find chatgpt.com has greater "intelligence" than codex... I was wondering if anyone has found a better solution than copying context to and from chatgpt.com... I really just want this intelligence to live on my laptop, instead of me being the intermediary.

变现:开源 CLI 与桌面工具免费,订阅制托管推理编排服务(按 token 或月费 10-20 美元),或卖 team 版多模型路由配置

MVP:做一个本地 CLI 与 MCP 服务:把任务规划阶段路由到高推理模型(如 ChatGPT 或 Claude 网页版或高配模型),产出计划后自动交给本地 codex 与 Claude Code 等执行 agent,用文件与剪贴板桥自动传递上下文,省去人工复制粘贴;2 周可做出 macOS 原型。

竞品:Cursor、Claude Code、aider、Continue.dev、OpenAI Codex CLI、LangChain/LangGraph 多模型编排

摘要:HN 用户反馈 chatgpt.com 的 Pro 级模型洞察力明显强于 codex agent,但工具调用能力远不如后者,目前只能手动复制上下文做中间人。他希望这种高智能能力直接落在本机上。帖子只有 5 分和 11 条评论,属于小众工作流诉求,没有明确付费表述,且多模型编排赛道已经非常拥挤。