PSI: Shared State as the Missing Layer for Coherent AI-Generated Instruments in Personal AI Agents
A shared-state layer called PSI lets independently generated AI mini-apps talk to each other — promising coherent personal software built from natural-language requests.
The Thesis
Today, when you ask an AI to generate a small tool — a habit tracker, a budget widget, a meeting summarizer — that tool lives in isolation. PSI proposes a 'shared-state bus' (a common memory layer that every generated module can both read from and write to), so that separately created tools automatically know what the others are doing. The practical effect is that your AI-generated habit tracker could automatically inform your AI-generated calendar widget without you wiring them together. The catch is significant: the entire evidence base is a single researcher using their own system for three weeks, which is not enough to draw broad conclusions.
Catalyst
Large language models can now generate functional UI components and small applications from plain-language prompts — a capability that barely existed two years ago. That shift creates a new coordination problem: how do you make dozens of auto-generated micro-tools behave like a coherent environment rather than a pile of disconnected apps? PSI is a direct response to that newly emerged problem.
What's New
Earlier personal productivity tools — think Notion, Obsidian, or even IFTTT-style automation — required manual integration work: the user or a developer had to explicitly connect data sources and define triggers. AI-generated tools inherit this isolation problem by default, because each generation call produces a self-contained artifact with no awareness of siblings. PSI proposes a publish-subscribe contract (each module announces its current state and offers 'write-back' hooks for other modules to act on) as a standard interface that any newly generated module can adopt automatically.
The Counter
The entire empirical basis here is one researcher, using a system they built themselves, for three weeks. That is an autobiographical case study, not a user study — and it cannot tell us whether PSI works for anyone other than its creator. The shared-state-bus idea is also not new: publish-subscribe architectures, event buses, and reactive state managers (Redux, MobX, RxJS) have existed in software engineering for decades, and the paper must explain why prior art failed in this context rather than simply proposing a new acronym for familiar patterns. There is no performance data, no comparison against a baseline coordination approach, and no evidence that 'later-generated instruments' integrate reliably at scale or across diverse user needs. Finally, the security and privacy implications of a 'shared personal-context bus' — where every generated module can read every other module's state — are not addressed, which is a serious gap for any system handling personal data.
Longs
- MSFT — Copilot Studio and Power Platform are the most direct incumbents that could adopt or be disrupted by this pattern
- AAPL — personal AI environment vision aligns with Apple Intelligence on-device ambitions
- PLTR — enterprise AI agents with shared context are a direct use case for Palantir's AIP orchestration layer
- BOTZ (robotics/AI ETF) — broad exposure to AI infrastructure underlying agentic personal computing
Shorts
- Zapier and Make.com — their core value is manual workflow integration; PSI-style automatic coordination could make that unnecessary
- No-code platforms (Bubble, Glide) — if AI generates and connects tools automatically, the market for point-and-click app builders shrinks
Enablers (Picks & Shovels)
- OpenAI function-calling and tool-use APIs — the mechanism by which generated modules could expose write-back affordances
- Local-first data frameworks such as Automerge and CRDT libraries — provide conflict-free shared state without a central server
- WebAssembly (WASM) sandboxing — lets generated modules run safely side-by-side in a shared environment
- Model Context Protocol (MCP) by Anthropic — an emerging open standard for giving AI agents structured access to external state, closely related to PSI's bus idea
Private Watchlist
- Replit — generates runnable apps from prompts and would need exactly this coordination layer
- Notion AI — closest incumbent personal productivity AI; shared state is a natural extension
- Dust.tt — enterprise AI agent orchestration with shared context
- Fixie AI — personal AI assistant infrastructure startup
Resources
The Paper
Personal AI tools can now be generated from natural-language requests, but they often remain isolated after creation. We present PSI, a shared-state architecture that turns independently generated modules into coherent instruments: persistent, connected, and chat-complementary artifacts accessible through both GUIs and a generic chat agent. By publishing current state and write-back affordances to a shared personal-context bus, modules enable cross-module reasoning and synchronized actions across interfaces. We study PSI through a three-week autobiographical deployment in a self-developed personal AI environment and show that later-generated instruments can be integrated automatically through the same contract. PSI identifies shared state as the missing systems layer that transforms AI-generated personal software from isolated apps into coherent personal computing environments.