Side by side
Agent Harness & Scaffold
A useful narrow distinction: the harness runs the interaction; the scaffold shapes how the model behaves.
This distinction is useful but not universal. Read each author’s boundary.
Swap columns ⇄| Dimension | Agent Harness | Scaffold |
|---|---|---|
| Purpose | Execute model–tool interactions. | Configure behavior and available context. [Hugging Face][Anthropic] |
| Role in the system | Runtime control. | Instructions and setup. [Hugging Face][Anthropic] |
| Inputs and outputs | Requests and results. | Instructions and tool descriptions presented to the model. [Hugging Face][Anthropic] |
| Persistence | Implementation-dependent. | Implementation-dependent. [Hugging Face][Anthropic] |
| Example | Dispatch a requested file read. | Provide project conventions. [Hugging Face][Anthropic] |
| Boundary caveats | Can include broader surrounding software. | Can overlap with harness. [Hugging Face][Anthropic] |
Evidence for this distinction
- Harness, Scaffold, and the AI Agent Terms Worth Getting Right ↗ (opens in new tab)Hugging Face · 2026-05-25
Relevant section: Scaffolding; Harness - Scaling Managed Agents: Decoupling the brain from the hands ↗ (opens in new tab)Anthropic · 2026-04-08
Relevant section: Virtualized components; Decouple the brain from the hands
Last editorial review: 2026-09-13.