SIGIL: Compiling Agent Skills into Typed Harnesses

2026-07-29

AI Summary

Current AI agents often skip critical steps when executing skills described in natural language due to the absence of structured control flow, resulting in unreliable and hard-to-verify behaviors. This work proposes a novel skill compilation paradigm that automatically translates natural language skills into typed, executable harnesses. By leveraging an intermediate representation—AG-IR—it decouples model cognition from program logic, thereby ensuring procedural completeness. The approach preserves the intuitive experience of natural language authoring while enabling model-agnostic, reliable execution. Experimental results demonstrate that the compiled harnesses successfully execute 86% of mandatory steps, achieve a 2.3× improvement in full-process completion rate, reduce token consumption to 0.58× of baseline methods, and maintain consistent performance across different model generations.

Abstract

AI-Integrated agents increasingly acquire capability from skills: prose procedure files loaded into a model's context and run by a tool-calling loop. A skill is described to the runtime but never encoded in it, so the model re-derives its control flow on every run and may skip mandated verification. Across 30 skills and two model generations, a prose agent performs only 56% of the steps its own skill mandates, while producing artifacts that pass output checks. The remedy is known: write a harness, in which the procedure is program structure. However, hand-writing harnesses is tedious and discards the authoring surface that made skills succeed. To address this limitation, we introduce Skill Compilation, realized in SIGIL, which compiles a prose skill into an executable harness. At its center is AG-IR, a typed agentic intermediate representation separating model-owned cognition from code-owned mechanism. Compiled harnesses perform 86% of mandated steps, complete the full procedure 2.3x as often, and require 0.58x the tokens. Notably, the guarantee is model-independent: the harness holds at 86% across two model generations while prose swings from 56% to 68%.

Problem

• agent skills
• prose procedures
• control flow
• verification
• executable harness

Innovation

• Skill Compilation
• AG-IR
• Typed Intermediate Representation
• Agent Harness
• Prose-to-Code Compilation