AgentScript is an approach to agent planning in which the model’s plan is itself a program. Rather than having the LLM describe what it intends to do and then act, it has the model emit a segment of plan code resembling JavaScript; that output is parsed into an abstract syntax tree, and a runtime executes it step by step.
The consequence its source draws is what the arrangement buys: the plan becomes explicit, reviewable, pausable and serialisable, which it says raises the agent’s interpretability and controllability. The source lists it among the representative implementations of Spec-driven development, alongside AI IDEs, CLI toolkits and research frameworks, and calls it a design for specification-driven agent behaviour.
Capabilities
The one source held here describes the mechanism in outline only: LLM emits plan code → parse to AST → runtime executes stepwise. It does not describe the language’s syntax, its tool-binding model, how a paused plan is resumed, or who publishes it.
Related Terms
- Spec-driven development — the practice the source lists this among the implementations of
- Checkpoint-and-Resume — a related idea; the source states that plans here are pausable and serialisable, but does not connect it to this term
- Human-in-the-loop — a related idea; the source treats human confirmation at key steps under spec-driven development generally rather than in its account of this design