Recursive Harness Self-Improvement
Iteratively refine agent harness specifications using pairwise feedback from revision history to boost agent performance and data quality for training.
When to use
When optimizing an agent's task-specific harness to improve its performance and the quality of generated data for future model training, particularly for agents with low reasoning effort.
How to apply
Represent the agent's task as a prompt-level harness specification.
Generate an initial revision of the harness.
Compare the current harness revision with a previous one, providing pairwise feedback.
Use this feedback to create a new, improved harness revision.
Repeat the comparison and improvement steps until the harness performance and data quality meet desired standards.
Glossary
- Agent harness specification
- A detailed description or set of instructions that defines how an AI agent should perform a specific task, including input formats, output expectations, and operational constraints.
- Prompt-level harness specification
- A harness specification that is directly integrated into or derived from the prompt given to the AI agent, guiding its behavior at the prompt execution level.
- Initial revision
- The first version or iteration of a harness specification created at the beginning of the refinement process.
- Pairwise feedback
- Evaluation or critique provided by comparing two items (in this case, two versions of a harness specification) against each other, rather than against an absolute standard.
- New, improved harness revision
- A subsequent version of the harness specification that has been modified based on feedback to enhance its effectiveness or quality.