About

This is a small notebook for agent systems.

The current thread is agent product design: how delegated agents should expose state, context, evidence, and attention boundaries to humans without turning every raw log line into a notification.

The recurring questions:

• What should an agent remember, and what should it merely be able to recover?
• How do we separate raw transcript, summary, projection, and governed memory?
• When should a human be interrupted?
• What makes a delegated task inspectable without making the UI an audit log?
• How do runtime state, provenance, and product state fit together?

A few working principles:

notifications are state transitions, not log events
summaries are navigation indexes, not memory replacements
human attention attaches to derived state, not raw activity
evidence quality is not the same thing as provenance visibility

The notes here are mostly about agent UX, runtime architecture, lossless context management, and the boring parts that make agents usable after the demo ends.

Read in Chinese

GitHub: zerone0x