DavidKPiano open-sources Jevspresso: XState constrains LLM agent actions in a coffee-bar demo
DavidKPiano · x · 2026-10-03
DavidKPiano (creator of XState) released Jevspresso, an open-source demo of a simulated espresso bar where the TypeSafe System One model Jev makes every barista decision, with an XState v6 state machine modeling what is physically possible.
The architecture inverts the usual agent pattern: instead of the LLM writing tool calls that the app validates afterward, the state machine defines the closed set of accepted events and Jev only picks from actions that are currently possible — it never writes an action itself. You can order in plain English or break the equipment to see how the model copes. Fully open source, runnable locally, live at jevspresso.dev.
More from coding & agent
- Vibe-coding a Japanese town with his daughter: Opus 5.5 outshines Astra — Merzmensch · 2026-10-04
- virattt's earnings KPI pipeline is now 1.8x faster, queryable in ~21s — virattt · 2026-10-04
- Meta paper: RL post-training hurts test-time scalability — the 'Sharpening Tax' — dair_ai · 2026-10-04
- AutoCompact trains agents to compact context themselves, +9.2 on SWE-bench Verified — omarsar0 · 2026-10-04
- Professor says AI grading passed his confidence test: TAs may never grade again — ipeirotis · 2026-10-04
- universal-modder: an open-source kit that lets AI coding agents mod almost any PC game, 2,300 stars in 3 days — alex_verem · 2026-10-04