New Tool Compiles AI Agent Instructions Into Reliable Code Harnesses

AI agents have begun relying on skills—step-by-step prose procedures fed into a model's context and executed through tool-calling loops. Yet a fundamental problem plagues this approach: because skills exist only as text, models must reinvent their control flow with each execution, often skipping required verification steps even as output checks pass.

The Problem with Prose Skills#

Testing across 30 skills and two model generations reveals the gap between instruction and execution. Agents following prose procedures complete only 56% of the mandated steps, despite producing artifacts that satisfy validation checks. This inconsistency persists across model versions, with different generations handling the same instructions differently (ranging from 56% to 68% compliance). The issue arises because skills are never truly encoded into the system—the model repeatedly re-derives how to execute them.

Compilation as a Solution#

The traditional remedy has been writing hand-crafted harnesses that embed procedure logic directly into code structure. However, this approach sacrifices the lightweight authoring process that made prose skills attractive in the first place. Enter SIGIL, a skill compiler that bridges these two worlds by automatically converting prose skill descriptions into executable harnesses while preserving the original authoring convenience.

How SIGIL Works#

At SIGIL's core lies AG-IR, a typed agentic intermediate representation that separates model-driven reasoning from code-driven execution mechanisms. This architecture lets developers define skills in prose while the compiler handles the translation into executable, structured code. The resulting harnesses enforce procedure compliance through code structure rather than relying on model behavior.

Results Across Models#

The improvements are substantial and notably consistent. Compiled harnesses achieve 86% step completion—a significant jump from 56%. Procedures execute to completion 2.3 times more often. Token usage drops to 0.58x the original level, reducing computational cost. Most strikingly, these performance gains hold steady across different model generations, whereas prose-based skills swing unpredictably between 56% and 68% compliance. The guarantee is model-independent: the compiled approach maintains its 86% floor regardless of which AI model runs it.