The Harness: Intelligence in Action

The Harness runs an agent in environments it can observe and act in. Give it a task in your own words: a language model plans and writes short programs, and the runtime executes them. You can watch, send a message, or take over.

Project website and demos · Source code · Documentation

Thinking while acting

An agent does not have to stop working between model calls. The Harness keeps programs, skills and reactive rules running while the model reasons about what to do next. Notes, tested skills and rules learned during a task can be used again in the same run.

From computers to interactive worlds

Computer use, games and robotics give the agent different ways to observe and act. The project website brings these settings together through demonstrations of continuous execution, learning while working and thinking while acting.

The cover illustrates these three settings; see the project demos for the actual runs. The public core starts small, with computer, browser and workspace packages installed as needed.

Get started

npm install -g @lmmslab/harness
harness

The Harness needs Python 3.11 or newer. The first launch asks which model to use and for its API key. Follow the installation guide and quickstart to run a task on your computer.