Hills: Formalize Agent Workflows by Separating Eval from Design Space
Only_Management_1010 · reddit · 2026-08-15
The author open-sourced hills, a CLI library to formalize AI agent optimization workflows by separating evaluation code from the design space.
Key Features:
- Version-controlled constraints (Hills): Evaluation criteria are fixed and versioned; any agent attempts to modify them are visible.
- HTML Journaling: Agents maintain an HTML log of findings (inspired by Karpathy's autoresearch) for human oversight.
- Trustless Monitoring: Allows running agents for days without manual tracing, provided the evals are robust and isolated.
More from coding & agent
- Yukon accelerates open research with intelligent agents — gajesh · 2026-08-15
- Nous Hermes introduces /loop command for recurring tasks — NousResearch · 2026-08-15
- TEMPO: Solving Long-Horizon Agent Training via Macro-Step Policy Optimization — teortaxesTex · 2026-08-15
- AGENTS.md vs. skills: How to steer a coding agent — rseroter · 2026-08-15
- Cursor Origin Launches for Agentic Workflows: 22 Commits per Second — omojumiller · 2026-08-15
- Alibaba's Metis agent reduces redundant tool calls from 98% to 2% — bendee983 · 2026-08-15