2 months ago
- LLM harnesses should be intuitive, transparent, lean, and error-resilient to reduce cognitive load on bots.
- Key principles include determinism in goal pursuit, minimal core prompts, and avoiding context limit issues.
- Leverage LLMs' coding knowledge by using familiar environments like Unix/Linux systems for efficiency.
- Harnesses need auditability, logging, and self-healing to handle failures at both LLM and harness levels.
- Adopt Unix philosophy: modular tools, interoperability, and text streams as universal interfaces.
- Organize data using a filesystem hierarchy (e.g., FHS) to help agents navigate and manage external data easily.
- Introduce a 'Kernel' event bus to monitor file changes and trigger LLM responses for real-time handling.
- Ambiance prototype includes default users like root, pai, and librarian for system tasks and interaction.
- Harnesses built on existing model knowledge (files, logs) outperform those requiring new training.