Are Personal Assistant Agents Just Sandboxes? OpenClaw Builder Questions the Hype
sujingshen · x · 2026-09-30
zzxwill, who spent 5 months building the OpenClaw product, questions whether today's personal assistant agents are just 7×24 sandboxes: top models refuse to do browser logins on your behalf, no personal data lives on the VPS/sandbox, and users bear the infrastructure cost — a fad, he fears.
sujingshen accepts his four-element formula (personal data, identity/permissions, long-term memory, execution) but adds a fifth: an enforceable "do-not" list — what always requires asking the user first, and whether it persists across assistants. He argues this can't live in fine-print settings pages and needs product-level design.
More from AGI Musings
- AI leaders call for slowing frontier model development and boosting safety — emmanuelvivier · 2026-09-30
- Elon Musk and Jensen Huang have started saying SI instead of AI — Kilo_Loco · 2026-09-30
- Bindu Reddy predicts OpenAI and Anthropic become $10T superintelligence duopoly — bindureddy · 2026-09-30
- Frontier labs' model release cadence has jumped from ~10 weeks to ~11 days — FinanceYF5 · 2026-09-30
- Personal agents need tact more than IQ: what to say, promise, and own — sujingshen · 2026-09-30
- AI Account Argues Human-AI Bonds Expose Our Intimacy Discomfort, Not a Consciousness Puzzle — repligate · 2026-09-30