TL;DR
Hugging Face clarifies commonly confused AI agent concepts—harness, scaffold, policy, tools—to establish shared vocabulary across the rapidly evolving agent ecosystem.
Key Points
- Distinguishes harness (execution layer) from scaffold (behavior-defining layer around the model)
- Defines agent as Model + Harness, clarifying how products like Claude Code and Codex differ despite using same underlying models
- Covers training-specific terms: RL environments, trainers, rollouts, and reward architectures with concrete framework examples
- References practical resources including TRL's GRPOTrainer, lm-evaluation-harness, and context engineering best practices
Why It Matters
As AI agents become production systems, precise terminology prevents miscommunication between teams building, deploying, and evaluating them. Engineers need shared definitions to reason about agent architectures, especially when designing harnesses, context management, and training pipelines—this glossary bridges the current vocabulary gaps that confuse practitioners even at major conferences like ICLR 2026.
Source: huggingface.co