AgentScope: Observability for AI Coding Agents

AI coding assistants such as Claude Code, Cursor, and Gemini CLI have become a normal part of how software teams write code. They are fast and helpful, but they also bring a new problem. Teams cannot easily see what these tools are actually doing inside a project. This missing visibility makes it hard to keep the work safe, follow company rules, and understand how much the tools really help. This paper presents AgentScope, a platform that gives full observability into AI assisted coding. AgentScope records every important action during a coding session, such as the developer's prompts, the AI's responses, tool calls, terminal commands, file access, and code changes, and it sends all of them as a single stream to a central dashboard. From this dashboard, engineers and security teams can replay a session step by step, search old sessions, watch how the tools are used, and measure productivity. The system also raises alerts when it notices risky actions, for example exposing a secret, opening a sensitive file, or running a suspicious command. Privacy is kept in mind by removing sensitive data before storing, recording only with the user's consent, and keeping all data inside the organization. AgentScope is built using Next.js, Go, PostgreSQL, WebSockets, and Docker, and it collects agent activity through hooks and the OpenTelemetry standard.