Agents inherit HF slop defaults: temp=0 creeping into rubric generation loops
kalomaze · x · 2026-09-13
kalomaze found his agent (astra) forcing temp=0 in a "generate rubric criteria candidates based on previous bests" loop, overriding a previous default for no explicable reason. His takeaway: there's now enough HuggingFace slop code on the internet for agents to pick up bad patterns as defaults — making explicit guardrails against them necessary. A concrete case of web training data polluting agent behavior.
More from coding & agent
- mitsuhiko on agentic coding: over-generating code poisons the agent's own context — mitsuhiko · 2026-09-13
- M3E Canvas turns browser sketches into vibe-coding prompts for Claude Code and Cursor — tom_doerr · 2026-09-13
- 28-Minute Free Video from Claude Code Engineer Covers CLAUDE.md, Memory Shortcuts and Parallel Sessions — nikola_mr64990 · 2026-09-13
- Build a domain-specific agent harness with PowerShell and PSAISuite — dfinke · 2026-09-13
- Mac MCP: open-source MCP server turns ordinary ChatGPT chats into a macOS control plane — bulutarkan · 2026-09-13
- Dad with zero coding experience ships an App Store app after 8 months building it entirely with Claude — BUYMEBONESTOORM · 2026-09-13