Definition
An Agent Harness is the comprehensive software runtime environment surrounding an AI model that supplies the context, tools, constraints, persistence layers, and feedback loops required to safely execute work.
Purpose
Bridges raw model reasoning with real-world infrastructure, preventing stochastic hallucination by enforcing testing, permission bounds, and error recovery.
Key Concepts
How It Works
Wraps the LLM in an execution loop: it injects relevant repository state, intercepts function calls, runs code in sandboxes, and feeds compiler errors back to the model.
Related Terms
Sources
- Birgitta Böckeler, AI Agent Harness, Thoughtworks (October 2026)
- Anthropic Engineering Guides