A developer named Scott has created Skillscript, a declarative and sandboxed language designed for tool orchestration [1].

This development addresses a core instability in how AI agents execute recurring tasks. By moving away from purely generative instructions, the language provides a versionable framework that ensures an agent performs the same action consistently across different sessions.

Scott developed the language specifically for use with his local agent, NanoClaw [1]. The project emerged from a need to standardize the way the agent handled daily routines. "I wanted my NanoClaw agent to run my morning brief the same way every day," Scott said.

Without a structured language, agents often rely on the model to determine the best path to a goal in real time. This process can lead to unpredictable results and unnecessary expenses. "Every session, it would re-figure out how to do this from scratch, drift a little, and cost tokens for what's," Scott said.

Skillscript functions as a sandboxed environment, which limits the potential for the agent to execute harmful or unintended commands on a host system [1]. Because the language is declarative, the user defines the desired end state rather than a sequence of volatile steps. This approach allows for better version control of agent skills, meaning a user can update a task definition without needing to retrain the agent or hope the prompt remains effective.

The project is currently in development as a tool for local automation [1]. By reducing the reliance on the model's reasoning for every repetitive step, Skillscript aims to lower the total token expenditure required to maintain a complex agent workflow.

“"I wanted my NanoClaw agent to run my morning brief the same way every day."”

The shift toward declarative languages for AI agents represents a move toward 'deterministic AI.' While large language models excel at creativity and reasoning, they struggle with precision and repeatability. By implementing a sandboxed orchestration layer, developers can create a hybrid system where the LLM handles the intent, but a rigid, versioned script handles the execution, thereby increasing reliability and reducing operational costs.