Keep it Simple, Stupid Agent Framework
KISS is a lightweight agent framework that implements a ReAct (Reasoning and Acting) loop for LLM agents. The framework provides:
- Trajectory Visualization: Web-based visualizer for viewing agent execution trajectories with modern UI
- Simple Architecture: Clean, minimal core that’s easy to understand and extend
- GEPA Integration: Genetic-Pareto prompt optimization for compound AI systems
- KISSEvolve Integration: Evolutionary algorithm discovery framework with LLM-guided mutation and crossover
- Model Agnostic: Support for multiple LLM providers (OpenAI, Anthropic, Gemini, Together AI, OpenRouter)
- Native Function Calling: Seamless tool integration using native function calling APIs (OpenAI, Anthropic, Gemini, Together AI, and OpenRouter)
- Docker Integration: Built-in Docker manager for running agents in isolated environments
- Trajectory Tracking: Automatic saving of agent execution trajectories
- Token Usage Tracking: Built-in token usage tracking with automatic context length detection and step counting
- Budget Tracking: Automatic cost tracking and budget monitoring across all agent runs
- Self-Evolution: Framework for agents to evolve and refine their prompts
- SWE-bench Dataset Support: Built-in support for downloading and working with SWE-bench Verified dataset
- RAG Support: Simple retrieval-augmented generation system with in-memory vector store
- Useful Agents: Pre-built utility agents including prompt refinement and general bash execution agents
- Multiprocessing Support: Utilities for parallel execution of functions using multiprocessing
Contributors
Koushik Sen