Show HN: Skillscript – A declarative, sandboxed language for tool orchestration

Skillscript is a novel, declarative programming language designed for AI agents to define and execute their workflows predictably and efficiently. The project addresses key challenges in current agent deployments, namely the repetitive, token-intensive re-derivation of tasks by LLMs, the lack of a persistent 'substrate' for agents to solidify their capabilities, and the inherent security risks of allowing agents to write arbitrary code in general-purpose languages like Python.

Skillscript aims to provide a solution by:

• Reducing Cost and Latency: By allowing agents to author fixed procedures that execute cheaply and quickly, rather than re-reasoning each time.
• Enhancing Auditability and Safety: The language is deliberately not Turing-complete, disallowing operations like arbitrary eval or subprocess.run, thereby limiting potential damage. Its declarative nature ensures human readability and static validation.
• Introducing a Connector Model: Skills interact with external tools, models (local or frontier), and data stores via a well-defined connector interface, centralizing security and configuration.
• Defining Skill Types: Distinguishing between 'Headless' (asynchronous output), 'Augmenting' (for agent context), and 'Template' (agent-executable prompt) skills, and the concepts of 'static' vs. 'dynamic' execution.
• Enabling Human Oversight: Agents can author skills, but humans are intended to review and approve them, especially for effectful operations, reinforcing a 'human-in-the-loop' trust model.

The core bet behind Skillscript is that the majority of agent-authored automation is 'dispatch-shaped' rather than 'computation-shaped.' By providing a structured, auditable, and safe language, Skillscript seeks to become the default substrate for agent-fired automation, analogous to how SQL became foundational for data access.