Agents are not Apps: Stop building interfaces, Start building coworkers
summary
An agent named Ella runs for five months over Telegram, handles RSVPs, builds an RSVP website, creates a budget tracker, and books four vendors for a wedding of 100 guests, running on about a $5 VPS and a ChatGPT subscription. Architectural advice: use event triggers (webhooks/cron), keep deterministic checks before handing off to agents for judgment, store memory in a file-system first, require verification (programmatic tests or human approval), and grant autonomy gradually.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:15Who I am and why I care about agents
- 2:42From tools to co-workers with AI
- 3:23Wedding planning agent demo
- 5:51What makes a co-worker
- 7:16AI co-workers examples
- 10:05PR summarization flow
- 11:29Outcomes over rituals
- 12:33Event triggers and memory
- 13:56Trust and autonomy
- 15:43Reliability over capability
- 17:32Verification and self-checking
- 18:38Product design for co-workers
- 20:00Trip concierge demo
- 21:24Unfinished product work
- 22:05Q&A
