SJTU-IPADS/SkVM

What it solves

SkVM addresses the problem of "skill portability" for LLM agents. Because different LLMs and agent harnesses (the software environments that run agents) have different capabilities and constraints, a skill designed for one model may not work efficiently or at all on another. SkVM provides a system to compile and optimize these skills so they can run reliably across heterogeneous models and environments.

How it works

SkVM operates as a compilation and runtime system using a four-part pipeline:

1. Profiling: It measures a specific model and harness combination against a set of primitive capabilities to create a capability profile.
2. AOT-Compilation: An Ahead-of-Time compiler rewrites the skill to align with the target model's profile.
3. JIT-Optimization: A Just-in-Time optimizer improves the skill's content and runtime speed. This can be done using synthetic tasks (looping through edit, rerun, and score) or by analyzing existing conversation logs to diagnose and fix issues.
4. Benchmark: It evaluates the performance of original, compiled, and optimized skills across various tasks and models.

Who it’s for

It is designed for developers and researchers building LLM agents who need their agent's skills to be portable across different model providers (e.g., Anthropic, OpenRouter) and different agent harnesses (e.g., OpenClaw, Hermes).

Highlights

• Cross-Model Portability: Enables skills to be compiled for specific target models and harnesses.
• Coded Optimization: Uses an internal coding agent to automatically refine skills based on execution results.
• Flexible Configuration: Supports both "managed" (fresh sandbox) and "native" (clones existing user config) adapter modes.
• Integrated Toolchain: Includes a CLI for profiling, compiling, and optimizing, along with a web UI for reviewing proposals.