Meet Luna https://luna.com.ai, the first generation of autonomous agents, which proved that agents can do real work. It also proved they broke. Unbounded self-coding that can touch any key. Everything is riding on raw LLM reasoning, so execution drifts. No way to see what actually happened.
Luna is the second generation. It still learns and improves itself, but on a scaffold built for trust:
- Deterministic workflows, the agent writes and refines, then runs on a real engine instead of re-reasoning every time.
- A one-way vault. Secrets resolve below the plugin layer and never enter the model's context.
- Rollback on every change, an audit trail on every action, and approval gates on anything sensitive.
- Everything else is a plugin: tools, channels, skills, isolated per agent.
Luna isn't a chatbot. You describe the goal, it builds the workflow, you correct it, and it runs reliably next time. A private Luna spins up in about a minute today.
There's no playbook yet for what a good agent is in this context. We're writing it with real use cases and real constraints. That's the job.