Every Software Just Became Agent-Native
Summary
On May 28, 2026, several key AI advancements were announced. HKUDS lab introduced CLI-Anything, a system that generates CLI harnesses for any software, making over 50 GUI-only applications like GIMP and Blender agent-controllable. This includes a CLI-Hub package manager and a 7-phase generation pipeline, ensuring native integration with major agent platforms. Paradigm and Tempo open-sourced Centaur, a Slack-native multiplayer agent system under Apache 2.0, offering isolated Kubernetes sandboxes, secure credential handling, and self-optimizing skills for team-based AI work. Microsoft Research unveiled SkillOpt, a novel approach to training agent skills by treating "SKILL.md" as a trainable parameter, achieving significant performance boosts, such as an 18.6-point average increase with Claude Code and +58.3% on Spreadsheet tasks, across 6 benchmarks and 7 models. These innovations are complemented by an ongoing inference price war, with DeepSeek V4-Pro now at \$0.87 per million output tokens.
Key takeaway
For AI Engineers building agentic workflows, these advancements significantly expand agent capabilities and deployment options. CLI-Anything allows you to integrate virtually any existing software into your agent's toolkit, while Centaur provides a robust, secure platform for deploying collaborative agents within team environments like Slack. Consider adopting SkillOpt to iteratively refine your agents' "SKILL.md" files, potentially boosting performance by over 18% on complex tasks. This shift towards agent-native software and trainable skills demands a re-evaluation of your current agent development and deployment strategies.
Key insights
AI agents are gaining universal software control and self-improvement capabilities through CLI harnesses and skill optimization.
Principles
- CLI serves as a universal interface for LLMs.
- Agent skills are trainable via textual edits.
- Multi-agent systems need secure, collaborative runtimes.
Method
CLI-Anything generates CLIs via a 7-phase pipeline. SkillOpt trains skills by having an optimizer model propose "SKILL.md" edits based on scored trajectories, accepting only performance-improving changes.
In practice
- Generate CLIs for GUI apps with CLI-Anything.
- Deploy Centaur for team-based, secure Slack agents.
- Optimize agent "SKILL.md" files using SkillOpt.
Topics
- AI Agents
- CLI Automation
- Skill Optimization
- Multi-agent Systems
- LLM Inference
- Open-Source Models
- Software Integration
Code references
Best for: AI Architect, Machine Learning Engineer, Entrepreneur, AI Engineer, MLOps Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by unwind ai.