Harness, Scaffold, and the AI Agent Terms Worth Getting Right

Summary

AI-agent terminology has outpaced shared definitions, leaving terms such as harness and scaffold blurred across frameworks and contexts. The glossary separates the model, which produces text but has no memory or execution loop, from scaffolding, which defines instructions, tools, output parsing, and context management, and the harness, which runs model calls, routes tool use, and decides when to stop. It presents an agent as a model plus the surrounding system that lets it act repeatedly, while distinguishing tools, skills, sub-agents, and orchestrators by their roles. For training, it defines the environment, trainer, rollout, and reward, including verifiable or learned scores and sparse or dense feedback. Because usage is not universal, the glossary offers a practical mental model rather than enforcing one vocabulary.

Takeaways

1. 1

   A model produces text but has no memory between calls or execution loop; scaffolding and a harness provide the instructions, tools, context management, and execution needed for agent behavior.

2. 2

   The harness runs model calls, handles tool calls, and decides when to stop, while an orchestrator coordinates multiple agents that may each run their own harness.

3. 3

   In training, a rollout records a complete agent run and its rewards, while the trainer scores many episodes and uses those results to update the model's weights.