A two-layer, seven-question framework for judging what makes a personal AI agent truly yours
sujingshen · x · 2026-10-30
Concluding an eight-part series reviewing dozens of companies, labs, open-source projects and standards bodies, the author proposes a personal framework: a personal agent is an AI system that serves one defined person long-term, knows and remembers them, and acts within their authorization.
Two layers, seven dimensions
- Layer 1 (what it can do): execution (end-to-end task complexity — the focus of national standards, Tsinghua AIR and CAICT gradings), personalization (memory, knowledge, proactive anticipation, plus whether memories are viewable/editable/deletable), and presence.
- Layer 2 (who it serves and on what basis): service object, authorization, interests, and allegiance — which truly determines whether an agent is "personal."
Key observations
- Product launches this year nearly all tout Layer 1; researchers, protocols and open-source focus on Layer 2.
- Capability means little without access: Doubao was rejected by 155 apps and Muse was blocked by Amazon.
- OSWorld completion rates show a big gap between advertised capability and actual performance.
More from AGI Musings
- Tim Ferriss: 90%+ of AI use cases are doing work that shouldn't be done at all — JeremyNguyenPhD · 2026-10-30
- "Questions are a commodity now," says investor Stewart Alsop III — StewartalsopIII · 2026-10-30
- "AI for Me, Not for Thee": A Rant on Developers' Double Standard — Entire-Picture3990 · 2026-10-30
- Scientists nostalgia: AI made molecule screening instant, but what did we lose? — charlieharris01 · 2026-10-30
- Austria blocks digital red-tape cut over notary revenue, mocking EU AI ambitions — mitsuhiko · 2026-10-30
- 'Biology is just a data problem?' One mitochondrial peptide shows why that's naive — iScienceLuvr · 2026-10-30