Nov Hands-On Agent Building

The front-end Router only chats and dispatches tasks, while back-end Workers work in isolation and in parallel, solving the bottleneck that a person can monitor at most 3 to 5 sessions.

Persistent Agents are triggered by GitHub or Feishu events. Permissions are split into authorization credentials and one-time work orders, and each order uses its own independent working directory.

Plan the follow-up commands and guard conditions once and cache them. When they match, execute directly, skipping repeated model calls, which can significantly save Tokens.

Compress operation logs in the background before reinjecting them into a new session, or distill successfully run tasks into a Skill so the Agent remembers the pitfalls it has stepped into.

China-based services can be connected to Agents through official MCP, Skill, or CLI, covering coffee ordering, ride booking, office collaboration, and content platforms.

Having one main Agent uniformly dispatch multiple CLIs such as Claude Code and Codex is simpler than switching between them one by one, dividing work by task.

Before starting, write a spec or technical plan. After writing, review the changes with a clean context to reduce the Agent veering off course and self-validating.

Before deploying a self-hosted Agent, close public port mappings, add authentication, and don't run containers as root, to prevent commands from being executed directly.