Releases·Tool Release
HEXIS compiles SKILL.md agent skills into extended finite state machines with LLM reasoning scoped inside states
HEXIS, a Python package released on GitHub alongside a paper, compiles agent skill documents (SKILL.md) into extended finite state machines. LLM reasoning is scoped inside states while transition guards over recorded variables enforce deterministic execution order, preventing steps from being skipped or reordered. It works with any OpenAI-compatible model and ships as MIT-licensed alpha software.
72 engagement1 sourceThu, Sep 17, 2026, 06:23 AM
AI Overview
- HEXIS, an alpha-stage MIT-licensed Python 3.11/3.12 package, compiles SKILL.md agent skills into extended finite state machines enforcing deterministic execution via transition guards
- Accompanies the paper 'Compiling Agent Skills into Extended Finite State Machines'; compilation works with any OpenAI-compatible model via --model and --base-url flags
Related
Developments·Paper
HEXIS compiles agent skills into extended finite state machines, beating Skill + ReAct by 16.1 points
A new arXiv paper introduces HEXIS, a system that compiles agent skills into extended finite state machines separating skill knowledge from control flow. Across four benchmarks and four executors, HEXIS improves success over Skill + ReAct by 16.1 percentage points on average and cuts execution tokens sharply with Qwen3.8-27B. The work addresses agents repeatedly inferring how to apply skills, which couples reasoning with control decisions and lets prescribed steps be skipped or misapplied.
A new arXiv paper introduces HEXIS, a system that compiles agent skills into extended finite state machines separating skill knowledge from control flow. Across four benchmarks and four executors, HEXIS improves success over Skill + ReAct by 16.1 percentage points on average and cuts execution tokens sharply with Qwen3.8-27B. The work addresses agents repeatedly inferring how to apply skills, which couples reasoning with control decisions and lets prescribed steps be skipped or misapplied.
0 engagement1 sourcearxivThu, Sep 24, 2026, 04:58 PM
Developments·Paper
Compiling SKILL.md into Jac raises agent step completion from 56-68% to 86%
A new approach compiles SKILL.md files into a compiled format called Jac, boosting agent task completion rates from 56-68% to 86% with 100% median compliance. This suggests that precompiling skill definitions can significantly improve agent reliability.
A new approach compiles SKILL.md files into a compiled format called Jac, boosting agent task completion rates from 56-68% to 86% with 100% median compliance. This suggests that precompiling skill definitions can significantly improve agent reliability.
1 engagement1 sourceredditFri, Aug 21, 2026, 11:05 AM
Social·Community
Implementing Agent Skills support: three loading decisions worth separating
A developer implementing skills support in a finance agent describes how 'load the skill' separates into three decisions: what the server loads, what the model sees, and when tools are available. The post doubles as practical guidance for building skills systems and links to a framework-agnostic guide with TypeScript/Python examples.
A developer implementing skills support in a finance agent describes how 'load the skill' separates into three decisions: what the server loads, what the model sees, and when tools are available. The post doubles as practical guidance for building skills systems and links to a framework-agnostic guide with TypeScript/Python examples.
4 engagement1 sourceredditWed, Sep 30, 2026, 01:10 PM
Products·launched
antigravity-skill-porter: ports Claude Code / Cursor SKILL.md skills to Google Antigravity
An open-source tool that converts community SKILL.md workflows written for Claude Code and Cursor so they run inside Google Antigravity. It deterministically maps Claude-specific tool calls (Bash, Read, Edit) to Antigravity equivalents and upgrades serial single-agent loops to parallel subagents. It solves the problem of skills breaking when moved between agent coding environments.
An open-source tool that converts community SKILL.md workflows written for Claude Code and Cursor so they run inside Google Antigravity. It deterministically maps Claude-specific tool calls (Bash, Read, Edit) to Antigravity equivalents and upgrades serial single-agent loops to parallel subagents. It solves the problem of skills breaking when moved between agent coding environments.
4 engagement1 sourceredditMon, Sep 7, 2026, 07:59 AM
Developments·Paper
SkillCorpus framework consolidates and evaluates open-source SKILL.md ecosystem for LLM agents
Researchers introduced SkillCorpus, a framework to aggregate, curate, match, and evaluate the fragmented open-source SKILL.md files that package reusable procedural knowledge for LLM agents. The work addresses the redundancy and uneven quality of these artifacts, aiming to consolidate them into a single usable corpus and assess their practical benefit on real-world agent tasks.
Researchers introduced SkillCorpus, a framework to aggregate, curate, match, and evaluate the fragmented open-source SKILL.md files that package reusable procedural knowledge for LLM agents. The work addresses the redundancy and uneven quality of these artifacts, aiming to consolidate them into a single usable corpus and assess their practical benefit on real-world agent tasks.
0 engagement1 sourcearxivFri, Jul 17, 2026, 01:55 AM