The working
notes.

Notes, lessons and experiments — from AI agents and automation to operations, tools and the work itself.

  1. Nine AI agents built on Claude Code run my marketing. What each one does, the apps it can use, how work moves between them through Notion, and what none of them can do without my approval.

  2. Let the agent read and draft on its own, and make it stop for a yes or no before money, client messages, calendar changes, deletions and sharing. The five actions to gate, a table, and a 10-person agency example.

  3. Four AI models wrote the same three blog intros. DeepSeek V4.1 Flash tied Claude Sonnet 5.5 at 3.93 out of 5 for about an eighth of the cost, Claude Haiku 5.5 was cheapest, and hidden reasoning made up most of every bill. Three prompts, one run each: an early signal for routine writing.

  4. Claude Haiku 5.5 finished my research-agent task in 35 seconds, ahead of DeepSeek V4.1 Flash and GPT-6 Luna. It wasn't the most accurate, and one task proves little.

  5. Upwork, OnlineJobs.ph, BELAY and seven more, with the details most lists skip: who can actually apply from where, what each site costs you, and the FTC's warning signs for fake VA jobs.

  6. Know some Zapier, Make or n8n? Here's the six-stage roadmap I'd follow next: LLMs, process mapping, APIs and webhooks, integrations, AI agents, and AI systems that run safely with human approval.

Project screenshot