Jensen Huang said it at NVIDIA GTC. Every company will need an agentic system. Not someday. Now.
So what can these agents actually do? Pretty much anything a human does on a computer. Coding. Debugging. Screen control. Device access. Scheduling. Multi-platform workflows. And the interesting part is when they fail at something, they learn from it and do better next time.
Two Systems I Studied
I have been studying two open source projects that lead this space.
OpenClaw is about being everywhere. One assistant across 23+ platforms. WhatsApp, Telegram, Slack, Discord, iMessage. Always on. Always reachable. 354K GitHub stars.
Hermes Agent is built on top of OpenClaw but the focus flipped. It learns your preferences. Adapts to your style. Gets smarter with every interaction. 57K GitHub stars.
Both are great in their own way. One is wide. The other is deep.
So I Started Building My Own
I called it AstraClaw. It is written in Python. Right now it is a small agent with a tool-calling loop. Nothing fancy yet. But the foundation is there.
What it can do so far:
- Main orchestrator loop
- Shell access with safety approval for risky commands
- File read and write tools
- File search by content and filename
- Full codebase explanation
- OS-aware commands (Linux and Windows)
- Session persistence as JSONL transcripts with token streaming
- Works with OpenAI and OpenRouter providers
What is coming next:
- Better memory
- Reusable skills
- MCP support
- Multi-device support
- Docker sandboxing
- More autonomous execution
Security First
Security has been a first class concern from day one. The system prompt is designed to detect the user’s OS. If your agent does not know the OS, file operations and shell commands will break in weird ways. Risky shell commands also need approval before they run. Small thing but it matters when you give an agent real access to your machine.
Follow the Build
The repo is live. If you want to follow the journey, a star on GitHub would mean a lot. Feedback, ideas, and critique are all welcome.