A Complete Breakdown of Agent Harness

The piece defines an AI agent harness as everything around the model: state, files, tools, execution environment, and orchestration.

It breaks the harness down into core pieces like filesystem access, Bash, sandboxing, tool use, and verification loops, explained from the model’s limitations.

LangChain’s Vivek Trivedy showed that improving only the harness can dramatically lift performance, not just the model.

In Terminal Bench 2.0, a coding agent moved from outside the Top 30 to the Top 5 after harness changes.