Agent control plane benchmark: Smalltalk-80 cascades hit 100% compliance vs 2.5% for Gherkin
RandalSchwartz · reddit · 2026-11-03
A dev recounts how their AI agent control plane accumulated five internal notations (Make, init-d, Graphviz DOT, Gherkin, Belnap tables) across 240+ agent tickets, then replaced all five with 1980 Xerox PARC Smalltalk-80 keyword cascades.
In a 360-trial benchmark under 16k tokens of distractor load:
- Prose + Gherkin + init-d: 2.5% compliance (3/120)
- Common Lisp S-expressions: 91.7% overall, dropping to 66.7% on flash-lite due to closing-paren drift
- Smalltalk-80 keyword cascades: 100.0% compliance (120/120, p < 0.0001)
The key insight: keyword selectors like state:invariant:onSuccess:onFailure: syntactically require every colon component, so omitting the sad path triggers an immediate compile-time MessageNotUnderstood error — deterministic grammar enforcing LLM output completeness beats fuzzy notations.
More from coding & agent
- Developer draws 'life of an animal cell' statechart in JS with Claude Opus and open-source statechart skill — DavidKPiano · 2026-11-03
- Developer finds Grok bot with Linear and Notion connectors a 'goated' capture-delegation system — malliktwts · 2026-11-03
- Devs flip on code bloat: from 'code is a liability' to praising AI slop spitting 50k unreviewed lines a day — rickasaurus · 2026-11-03
- Founder playbook for 2027: your moat is proprietary AI knowledge, marketing is what's left for you — alexmacgregor__ · 2026-11-03
- Three years building an AI RPG taught this dev to test what changed, not what AI said — tschilpi · 2026-11-03
- Dev builds "Lie Room" for Claude, Codex and Grok — Claude lied best, Codex got eliminated — Severe-Tank-5252 · 2026-11-03