5 个云端助理 + 便宜模型组一队
一份 OpenAI Dots 指南拆解如何用 5 个云端助理加便宜模型组成团队，含 6 步架构与成本、上下文隔离等细节。
5 CLOUD AGENTS + A MODEL 80% CHEAPER = A FULL TEAM AT THE TOKEN PRICE OF ONE AGENT the openai dots guide behind this pulled 328,922 views. 505 bookmarks its whole architecture sits on one page below. 6 steps. 6 exact numbers most setups are one chat that stops when the laptop shuts here each agent gets its own cloud machine: browser, terminal, files, memory, scheduler 6 steps: 1. hire > a new agent passes 5 tests before it exists: one domain, own sources, own style, a hard approval stop, a trigger > a vague general-helper role fails all 5 2. connect > 5 actions from opening the remote browser to handing control back > you log in by hand, 2fa included. the session stays on the agent's machine > 4,000+ apps on the integrations list 3. delegate > 1 lead splits the goal across a 5-agent roster and merges the results > each worker sees only its slice of context 4. clock > weekdays at 08:30 a routine builds a 4-section brief and posts it to slack > about 260 runs a year, nobody touching it > reading and drafting are free. outside changes wait for a human 5. models > sol costs one fifth of astra per token > five sol agents cost about what one astra agent does > faster astra tier: up to 8x quicker for coding 6. ship > a bug report becomes a reviewed pull request through a 5-stage loop in a cloud sandbox 25 5 tests x 5 agents. that many pass/fail checks before day one net: one chat that goes quiet with the laptop lid -> a lead and four specialists on their own cloud machines, running on a schedule and waiting for a human before anything changes outside state what isn't confirmed: > the author says the model mix cuts spend by over 70%. a claim, no independent number > the 1.05m-token sol window, roughly 790,000 english words in one prompt, is claimed the same way > dots launched for pro, business premium and enterprise plans. check yours build this once and your work moves from typing prompts to signing approvals architecture and all 6 steps on one page ↓
这条是引用着另一条写的：x.com/mojunel/status/2106042827679408577