A new open-source SDK from ClioAI targets AI agents for knowledge work like research and analysis. Unlike frameworks focused on code execution, this tool treats knowledge tasks as engineering problems with structured verification loops.

The SDK introduces an orchestrator pattern that coordinates subagents, web search, and file operations. It uses a unique approach where rubrics are generated separately from task execution, preventing agents from gaming the evaluation system. This creates a verifiable feedback loop for tasks that traditionally lack clear success metrics.

Key innovations include explore mode for generating multiple distinct approaches to creative problems, and checkpointing for resuming multi-agent workflows. The verification step acts as a reward function for reinforcement learning, making it valuable for training models on knowledge work. MIT licensed and available on GitHub, the SDK supports multiple AI providers including Gemini, OpenAI, and Anthropic.