Open-Source SDK for AI Knowledge Work

ClioAI has released the kw-sdk, an open-source Software Development Kit designed to bring structure to AI agents performing tasks in knowledge work sectors such as research, analysis, strategy, and writing. Unlike traditional code frameworks, which have natural verification signals through testing, knowledge work requires structured approaches for task verification and evaluation.

This SDK facilitates task execution through an orchestrator that manages subagents, web searches, code execution, and file I/O, followed by a verification step that checks the work against a pre-defined rubric, which remains hidden from the executor. This method simulates a reward function, ideal for reinforcement learning models applied to knowledge tasks.

Key to the framework are the concepts of Task Briefing, Rubric Creation, Execution, Verification, and Submission. The 'Explore mode' is particularly noteworthy; it allows the generation of multiple distinct approaches to a problem, providing insights into the solution space and highlighting gaps or trade-offs.

Another notable feature is Checkpointing, which aids in managing multi-agent systems by allowing you to save logical states during execution for further exploration or re-execution. This is beneficial when needing to rerun or iterate on specific stages of a process after initial tasks like searches have been conducted.

The SDK also supports remote execution environments. You can run commands within contexts such as local environments, Docker containers, or sandboxed browsers, making it versatile for different operational needs. This approach turns code execution into a protocol, giving flexibility in integration.

For installation, it is recommended to use Python and install directly from the GitHub repository: pip install git+https://github.com/ClioAI/kw-sdk.git. You can also add it to your pyproject.toml for development dependencies. An .env file is recommended for API keys like Gemini, OpenAI, and Anthropic.