Jev's structured-output model impresses, Yegge argues fences beat sandboxes for agents
njyx · x · 2026-09-19
Steampunk AI's weekly roundup hits three threads:
- Typesafe's System One model Jev: no chat, just structured input to structured output — effectively a fast, configurable super-classifier for high-bandwidth tasks with limited output classes. The author has hands-on access and is impressed; the name nods to Kahneman's System 1/System 2.
- Steve Yegge's "fences, not sandboxes": rather than squeezing agents into tighter sandboxes to prevent escape, let agent collectives evolve coordination mechanisms within defined boundaries — the author sees this as where agentic systems land over the next few years.
- Also: Gemini joins the hacking party, GitHub workflows show strain with replacements emerging, "slop grenades" is the term of the month, and Anthropic picks Accenture for AI audits.
More from coding & agent
- Cursor Ultra perks via SuperGrok Heavy quietly cut 75%: $400 to $100 quota — mazzaTalk · 2026-09-19
- Are 300-Sub-Agent Swarms Overkill? Kimi K2.6 Sparks Agent Architecture Debate — Relevant-Potential17 · 2026-09-19
- Community Showcases 350+ JEV Demos: Semantic Ctrl+F, Log-Mining PR Bots — vista8 · 2026-09-19
- Are MCP apps actually getting used? A dev rethinks SaaS form factors — The_Gluxer · 2026-09-19
- Mnemos engine builds synaptic-like agent memory via resonant engrams — RileyRalmuto · 2026-09-19
- Codex's built-in browser now supports extensions and cookie import — vista8 · 2026-09-19