The idea in one paragraph.
Personal AI stacks fuse prompts, tools, memory, and runtime around one cloud model, so swapping in a local model breaks everything at once. Decompose the stack into five typed primitives and optimise them jointly.
THE MACHINERY
Typed spec
The whole stack as one declarative, editable configuration object.
Five primitives
Intelligence, Engine, Agents, Tools & Memory, Learning - each independently editable.
LLM-guided spec search
A frontier model diagnoses failures and proposes coordinated edits; a held-out gate accepts only non-regressing ones.
Search-time cloud, inference-time local
Frontier models improve the spec; the resulting spec runs entirely on-device.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Swapping Claude Opus 4.6 for Qwen3.5-9B drops accuracy by 25-39 pp across personal AI tasks like PinchBench and GAIA.
Inspect a source excerpt
swapping Claude Opus 4.6 for Qwen3.5-9B drops accuracy by 25-39 pp across personal AI tasks like PinchBench and GAIARead the passage in context on arXiv ↗
State-of-the-art prompt optimisers close just 5 pp of the local-cloud gap on their own.
Optimised on-device specs match or exceed cloud accuracy on 4 of 8 benchmarks and land within 3.2 pp of the best cloud baseline on average.
They reduce marginal API cost by ~800x and end-to-end latency by 4x.
Expanding the editable set from prompts alone to all four primitives adds 5.5-16.5 pp.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Search-time still depends on frontier cloud models, so privacy holds at inference but not during optimisation.
- Evaluated on eight personal-AI benchmarks; the spec abstraction is not shown to transfer to other agent domains.
Why it belongs in this atlas.
The strongest empirical argument in the corpus for why harness design is a research problem: optimising one primitive plateaus at 5 pp, optimising four jointly closes 13-32 pp. The bottleneck was never the prompt alone.