Glossary
What is an agent harness?
In short
An agent harness is the software around a language model: a program that gives it tools, memory, instructions and rules, runs it in a loop and checks the result.
AI agents · harness
MMOOZZGG.com editorsChecked: 1 min read
For a child
A horse has a harness: straps, reins, a saddle. The horse is strong, but without a harness it will not pull a cart where you need it to go. The English word harness means exactly this. People use it for everything they put on a smart program: hints, tools, rules. Thanks to this, the program does not just chat; it does the job and stays on the road.
In plain words
A language model on its own can do one thing: take text and write what comes next. To make it work as an agent, people build a program around it. This program passes the model an instruction and the history, gives it tools (search, files, a CRM), runs the calls the model asks for, compresses a long conversation and decides when to stop. This wrapper is called a harness. The same model works very differently in different harnesses. There is a second meaning too: a set of programs for testing models on benchmarks, for example the open project for evaluating language models from EleutherAI.
For business
A company chooses not only a model but also the harness around it, and the quality of an agent depends on the harness no less than on the model. A good harness keeps a log of actions, limits rights, saves progress between sessions and checks results with tests. In November 2025, Anthropic described how an agent working for a long time is helped by a progress file, a task list with statuses and a browser check of each function: without them the agent tried to do everything at once or declared the work finished too early. Cost is made of development, integrations and the number of model calls per task. Changing the model inside a ready harness is usually easier than rewriting the harness, so test the architecture in a pilot, not in a presentation.
For a specialist
An agent harness (also called a scaffold) is the execution layer around the model: the system prompt, tool descriptions and execution, context management (compression, or compaction), memory between sessions, permissions and stop criteria. Anthropic calls the Claude Agent SDK a general-purpose agent harness. Its article Effective harnesses for long-running agents (26 November 2025) proposes, for work across many context windows, two agents with a shared harness: an initializer prepares an init.sh script, a progress file and the first git commit, and a working agent takes one function from a JSON list in each session and checks it with browser automation. The second meaning is an evaluation harness, such as lm-evaluation-harness from EleutherAI, which collects more than 60 academic benchmarks with hundreds of subtasks; the Open LLM Leaderboard on Hugging Face ran on it. In both cases the scores depend on the harness: the same weights give different results with different prompts, tools and scoring rules.
Example
A team connects a model to a harness for processing incoming invoices: the program picks up emails, gives the model access to the supplier directory, saves each step to a log and sends only disputed documents to the accountant.
Related terms
Questions and answers
How is a harness different from a model?
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map