Personal Life Automation Agent Stack: OpenAI Codex + Google Suite
TL;DR · AI Summary
Nicolas Bustamante shares his personal life automation agent stack: powered by OpenAI Codex, integrated with Google tools and Drive as data source, orchestrated via Skills for cross-app workflows; key decisions include using Drive as Source of Truth, contact CSV as hub, and implementing approval gates + feedback loops for reliability.
Key Takeaways
- Agent’s core capability is cross-app orchestration—not Q&A; e.g., intro email wo
- Drive + Markdown/CSV is agent-friendly data format—superior to Notion’s nested p
- Tool priority: API/CLI > local files > browser automation > screen/UI automation
Outline
Jump quickly between sections.
An agent is an orchestrator—not a question-answerer; users give 10-second requests, agents handle ~20-minute context-switching tasks.
Intro email requires 5 tools (WhatsApp→Gmail→web search→draft→approve→send); license renewal shows administrative continuity—same info in multiple systems.
Migrated knowledge from Notion to Google Drive intentionally—Drive + Markdown/CSV supports search, diff, editability, file-ID referencing, better than Notion for agent tool paths.
Tool priority: API/CLI > local files > browser automation > UI automation; Skills are iterative operation manuals (e.g., inbox-zero includes archiving logic, reply suggestions, signature policy).
High-stakes actions require 'read context → draft → show → approve → execute → confirm'; failures update Skills or write to AGENTS.md for continuous improvement.
'What did I miss?' scanning inbox is highest-value flow; checklist includes: install runtime, centralize data, cautious permissions, write rules (AGENTS.md), build & iterate Skills.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- 个人生活自动化 Agent 工具栈
- 核心定位
- 跨 App 编排执行者（非问答）
- 降低上下文切换成本（人工20分钟→用户10秒）
- 关键技术组件
- Agent 引擎：OpenAI Codex
- 工具层：Google全家桶+WhatsApp/Telegram+浏览器自动化
- 数据层：Drive（Source of Truth）+ CSV/Markdown
- 架构决策
- 数据形态：纯文本/表格/JSON命令 → Agent友好
- 工具优先级：API/CLI > 文件 > 浏览器 > UI
- Skill 设计：操作手册式可迭代规则
- 信任与反馈机制
- 批准门控：低风险直接执行，高风险四步走
- 反馈闭环：工具失败→修工具；判断错→更新Skill；遗忘偏好→写入AGENTS.md
- 杀手级工作流
- What did I miss?：定期扫描多源收件箱
- 收件箱分诊：识别urgent/stale/ignore/建日历/查文档
Highlights
Key sentences worth saving and sharing.
Notion is human-friendly but agent-unfriendly—nested pages, DB properties, permissions, native UI structure block automation; Drive + Markdown/CSV is truly agent-ready.
Contact CSV is called 'best investment'—it's the cross-channel lookup hub for phone/email/LinkedIn relationships.
Tool priority: API/CLI > local files > browser automation > screen automation; browser automation is only a safety net, not the main path—reliability capped by tool surface.
+ Tools: Google Suite, WhatsApp, Telegram, browser automation, etc. + Data: Google Drive, Notion, AGENTS.md, etc. + Skills: inbox-zero, contacts, etc. == Personal Life Automation Agent Tool Stack
@nicbstme proposed two typical workflows:
- Introduction Email (a “glue job” across 5 tools)
Friend’s WhatsApp request → search WhatsApp/Gmail for email → web lookup on company funding → draft introduction letter → await approval → send email → WhatsApp update on completion. Human effort: ~20 minutes, heavy context switching; user-side demand: ~10 seconds. The agent orchestrates cross-app interactions—not answering questions.
- License Plate Update (administrative continuity)
Upload photo to agent → update Markdown vehicle record in Drive → preserve VIN, insurance fields, etc. → upload back to Drive → optionally sync via browser automation to non-API systems like FasTrak, parking apps, or insurance portals. This illustrates administrative continuity: the same data remains consistent across multiple places—not one-off Q&A.
Key architectural decision: Drive as Source of Truth Nicolas deliberately migrated knowledge from Notion to Google Drive—pragmatic reasons:
- Notion is human-friendly but agent-unfriendly (nested pages, database properties, permissions, native UI structure).
- Drive + Markdown/CSV: searchable, diffable, editable, uploadable, referenceable via file ID.
gogcliprovides a unified CLI interface (Gmail, Drive, Calendar, Docs, Sheets, etc.).
Knowledge organization shouldn’t serve only human UIs—it should be structured for agent tool paths. Stable file IDs, plain text, tabular formats, JSON-returning commands—these are the data forms agents prefer.
Contact CSV (phone, email, LinkedIn, etc.) is called “one of the best investments,” as it serves as the hub for cross-channel lookups.
Tool priority (reliability tier): API / CLI > local files > browser automation > screen/UI automation Agent reliability is bounded by tool surface area. gog gmail messages list --json is more stable, retryable, and reason-able than asking the model to click around on a webpage. Browser and screen automation are fallbacks—not primary paths.
Skills: Agent’s “habits” and “taste” A skill isn’t fancy architecture—it’s an iterative operational manual. Take *inbox-zero* as an example:
- List inbox → distinguish auto-archived vs. manually reviewed items
- Highlight important emails, quote original text, suggest archiving or replying
- Draft first, wait for explicit approval before sending
- Preserve all recipients, keep replies concise, avoid unsolicited calls, sign off with “Nicolas”
Without skills, every preference must be re-prompted each time. With them, just say “run inbox zero.” Personalization of your agent comes from accumulated operational taste—not cute voice.
Feedback loop:
- Tool failure → fix tool or add guardrail
- Misjudgment → update skill
- Forgotten preferences → log to memory / AGENTS.md
- Repetitive workflows → improve system-wide through compound design
Approval gateways: Trust levels define product design Nicolas explicitly rejects “YOLO fully automated”:
- Low-stakes: direct execution allowed (e.g., “Tell Hugo I’m in Seattle next week”)
- High-stakes: mandatory steps: read context → draft → display → await approval → execute → confirm
The line between useful and scary lies in *when* you ask the human.
“Killer” workflow: *What did I miss?* More valuable than single-email automation is inbox triage for daily life:
- Every few hours, ask: “What did I miss?” → Agent scans WhatsApp, Telegram, Gmail, SMS, Calendar, Drive changes → summarizes: who needs reply, what’s urgent, what’s stale, what can be ignored, what should go in calendar, what requires docs.
Features: high-context, repetitive, cross-tool, full of small decisions—humans hate doing the first scan; agents excel at it; final judgment stays with humans.
Reproducible checklist (Nicolas’ path):
- Install agent runtime + channel CLI/connectors
- Centralize data: Drive as truth source, contact CSV, searchable documents
- Careful authorization: Full Disk Access, Screen Recording, Accessibility—must pair with peer-level approval gates
- Write operating rules (AGENTS.md): draft-before-send, tool routing, privacy boundaries
- Build skills for repeatable processes—and update after every misstep
Quote
Nicolas Bustamante
@nicbstme May 30
My Agent Stack For Automating My Personal Life
My agent manages my emails, SMS, WhatsApp, Telegram, and pretty much everything to automate my personal life. People keep asking me how I use agents in real life. I mean the actual boring things that...