A LoopScript is the artifact Structured Agentic Software Engineering (SASE) (SASE) proposes for defining how agents execute a task, introduced in Agentic Software Engineering: Foundational Pillars and a Research Roadmap as the product of Agentic Loop Engineering (ALE) (ALE). The paper motivates it by noting that agents cannot infer the “stakes” of a task on their own — they may “overthink” a simple request or under-deliver on a critical one — so a coach needs an explicit way to communicate the required level of rigor.
Usage
A LoopScript can specify task decomposition and parallelization (assigning one BriefingScript to multiple agents or a heterogeneous team of specialized agents, enabling N-version programming — e.g. a developer resolving seven tickets triggering 28 parallel pull requests, four per ticket), workflow strategy (granting full autonomy for a simple bug fix while enforcing a strict, multi-stage review process for a critical security patch), and evidence-based acceptance criteria (defining the structure of the final deliverable, a Merge-Readiness Pack (MRP)). The paper describes it as a living document a coach can dynamically adjust, for instance allocating more agents to a promising path or adding a review checkpoint if early results look uncertain, and frames it as a direct descendant of DevOps practices — declarative pipelines, infrastructure-as-code, and observability. It notes that some frontier coding agents already show elements of this: Google’s Google Jules has included a planning step from its inception, while Anthropic’s Claude Code only recently added an on-demand planning mode in which the agent generates a plan and awaits human review before proceeding.
Related Terms
Agentic Loop Engineering (ALE), BriefingScript, Merge-Readiness Pack (MRP)