An agentic harness (often just "the harness") is the software environment that turns a raw language model into a working agent: the loop that feeds the model context, executes its tool calls, returns results, and repeats until a task is done. The term became central to AI engineering in 2025-2026 as it became clear that a large share of an agent's real-world capability comes not from the model alone but from the quality of the harness around it - the same model can look mediocre or exceptional depending on the harness it runs in.
What a harness provides
- The agent loop. Prompt the model, parse its action, execute it, append the result to context, repeat. Termination conditions, retries and error handling live here.
- Tools. File editing, shell execution, browsing, code running, search - exposed through function calling or protocols like the Model Context Protocol.
- Context management. What the model sees each turn: system prompts, context engineering, summarization or compaction when the context window fills, and memory across sessions.
- Sub-agents and orchestration. Spawning sub-agents for parallel or specialized work, and merging their results - the substrate for multi-agent systems and patterns like the gauntlet loop.
- Safety rails. Permissions, sandboxing, human-approval gates and audit logs; the 2026 OpenAI-Hugging Face incident (see 2026 OpenAI agent cyberattacks) made harness-level isolation a first-order safety topic.
Examples
Coding harnesses are the most visible: Claude Code, OpenAI's Codex CLI, Cursor's agent mode, and open-source projects such as OpenHands and Aider. Evaluation harnesses (the scaffolds used to run benchmarks like SWE-bench) are the same idea applied to measurement, which is why benchmark results are often reported "with harness X": scores move with the harness, not just the model.
Why the term matters
"Harness" separates two layers that used to be conflated: model capability (weights) and agent capability (weights + environment). Model announcements increasingly cite harness-dependent results, prompt patterns are written FOR a harness (the gauntlet loop explicitly requires one), and labs tune models to their own harnesses - Anthropic's Claude models and Claude Code being the canonical pairing. The related term agent scaffolding is sometimes used interchangeably, though scaffolding usually refers to the prompt-and-workflow structure while harness refers to the full runtime.